Computer Science Department
University of Southern California
Los Angeles, CA 90089
1The definition of a node is architecture dependent. While in a multi-disk architecture a node may correspond to a disk drive, in a
shared-nothing architecture a node may consist of a CPU, one disk drive and some random accessmemory.
Data Declustering in PADMA: A PArallel Database MAnager
Jaideep Srivastava, Thomas M. Niccum, Bhaskar Himatsingka
Computer Science Department
University of Minnesota
Parallel processing of database operations was first addressed by the database machine community, where the
focus was on designing special-purpose hardware . However, the cost of building special-purpose hardware
is high, and most of the proposals were never realized . The eighties saw the emergence of very powerful and
scalable commercial massively parallel processors (MPPs), with extremely attractive price/performance ratios,
e.g. nCUBE, Intel’s iPSC and Paragon, KSR-1, and ThinkingMachines CM-2 and CM-5. Also, with very high
speed communication switches becoming commercially available, e.g. ATM and Fiber Channel, and advancements
in operating system technology to make communication cheaper, e.g. Active Messages , a network
of workstations (NOWs)  can be configured to provide the performance and price/performance of scalable
parallel machines. Thus, while special-purpose hardware design for databases did not succeed, use of MPPs or
NOWs for building parallel databases is an extremely promising and active research area.
The past few years have seen growing activity in the area of parallel databases. The relational data model,
whose set-oriented non-procedural nature provides opportunities for massive parallelization, has been found especially
suitable . A number of parallel database projects have been started in academia [3, 7] and industry
[1, 21] and products such as Parallel Oracle, Tandem Himalaya, Sybase Navigator, Teradata parallel database on
NCR DBC/1012, etc. are available in the market. Applications targeted range from transaction processing to
In this paper we provide a brief overview of the ongoing PArallel Database MAnager (PADMA) project at
the University of Minnesota and summarize specific results obtained in the area of data declustering. We finally
outline the project status and future directions.
2 Data Declustering : An Overview
Record-oriented data can be visualized as points in multi-dimensional space, with the co-ordinate of a point on
an axis being the value of the corresponding attribute of the record it represents. The declustering problem thus,
is deciding how to partition the entire data space into subspaces, which may be overlapping or non-overlapping,
and then deciding how to allocate data subspaces to disks. In general, it is possible to have multiple subspaces
allocated to the same disk, as well as a subspace allocated to multiple disks (replication). A data point (record) is
stored on the disk(s) to which the subspace containing it is allocated. Several declustering techniques have been
proposed in the literature, and good surveys are provided in   . A major class is of single-attribute
declustering methods, where the space partitioning is based on a single attribute. Examples are  , where
themost frequently queried attribute is used for declustering. Another classification of declusteringmethods can
be based on whether the partitioning of the data space is done in terms of regular grids, e.g. grid-file  type
partitioning, or irregular shapes . Though the question of whether regular or irregular partitions are better is
by no means settled, our focus has been on finding good declusteringmethods for regular grid-based partitioning
or cartesian product files All these methods and the class of methods we are studying work well for read-only
databases. They also handle well behaved updates, though further study is required in this area. Components of
the declustering problem are:
Creating the grid partitions, i.e. dividing the complete data space into regular sub-spaces.
Assigning individual sub-spaces to disk(s) i.e. disk allocation. Since we do not consider replication, this
is equivalent to finding a mapping which maps each sub-space to a unique disk.
Various deterministicmethods have been studied  for creating a grid partitioning of the data space. However,
these techniques are applicable only to small data files. Hence, statistical sampling based approaches become
extremely important. We have studied sampling based approaches for creating the grid partitions, and the
techniques have been shown to have very good partitioning properties. Details of these techniques are provided
in . In the following sections, we assume that the grid partitioning has been created. We thus use the terms
data declustering and disk allocation interchangeably.
2.1 Problem Definition
We now define some terminology which is used through out this paper. These definitions are similar to those
used by Faloutsos et al .
M Number of disks
d Number of attributes
Di Domain of i -th attribute
di Number of Intervals of the
domain of the i -th attribute
diskOf() Function that maps bucket-ids to disks
Definition 1 [Cartesian Product File] LetDi denote the domain of the ith attribute of a d -attribute file. Let each
Di be partitioned into di disjoint intervals Ii0; Ii1; :::; Iidi1. We call F a cartesian product file if all records in
partition I1i1I2i2:::Idid, where each Ijij 2 fIj0; Ij1; :::; Ijdj1g, lie in the same unique bucket(disk block).
The bucket b I1i1 I2i2 ::: Idid is denoted by < i1; i2; :::; id >.
Definition 2 [RangeQuery] Arange queryQ = [L1; U1)[L2; U2):::[Ld; Ud); Li; Ui 2 Di, is represented
as a d -tuple ([L1; U1); [L2; U2); :::; [Ld; Ud)). Here [Li; Ui) is the range on the ith attribute. Records that satisfy
this query must be points that lie in the d -dimensional box [L1; U1) [L2; U2) ::: [Ld; Ud).
Definition 3 [Partial Match Query] A partial match query Q is a range query such that
f(9i)[Li; Ui) Di] ^ [(8j 2 f1; 2; :::; dg)(j 6= i)(Lj = Uj )]g.
Definition 4 [Point Query] A point query Q is a range query such that
[(8i 2 f1; 2; :::; dg)(Li = Ui)].
Definition 5 [Length of Query] Let Q = ([L1; U1); [L2; L3); ; [Ld; Ud)) be a range query. The length of
Q on dimension i is the number of intervals intersecting [Li; Ui) on dimension i.
Definition 6 [Response Time] The response time of a query is defined as : max (N0;N1; :::;NM1) where
Ni(0 i M 1) is the number of qualifying buckets on disk i, for the query.
Since I/O is themajor bottleneck in query processing, it is desirable that I/O be parallelized as far as possible.
This becomes particularly important for a query which occurs frequently in a database system. The following
definition of query optimality gives the maximum possible I/O parallelization feasible for a query.
Definition 7 [Query Optimality]An allocationmethod onM disks is query optimal for a queryQif the response
time of query Q is d PM
e, where P is the total number of qualifying buckets for query Q.
Definition 8 [Strict Optimality] An allocation method is strictly optimal if it is query optimal for all possible
queries. It is strictly optimal for partialmatch queries if it is query optimal for all possible partial match queries.
It is strictly optimal for range queries if it is query optimal for all possible range queries.
2.2 Survey of Grid Based Declustering Techniques
We now provide a brief overview of the multi-attribute grid-based declustering approaches. These descriptions
are only to recapitulate their salient points. Detailed descriptions exist in the respective papers.
Figure 1 provides an example of how each of these techniques allocates a 2 dimensional grid, with 8 intervals
on each dimension, onto 4 disks.
1. Disk Modulo (DM) / Coordinate Modulo Declustering (CMD) The disk modulo method by Du and
Sobolewski  and coordinate modulo declustering by Li et al  are similar approaches. A bucket <
i1; i2; :::; ik > is assigned to the disk unit diskOf(i1; i2; :::; ik) = (i1; i2; :::; ik) mod M. Variations of this
method include the Generalized Disk Modulo allocation method .
2. Field-wise Exclusive-or (FX) This allocationmethod was proposed by Kim and Pramanik  with efficient
partial match retrieval in mind. The main idea behind this approach is the use of bitwise exclusive or
) on the binary values of a bucket-id. If < i1; i2; :::; ik > is a bucket-id then the FX method
allocates it to disk unit diskOf(i1; i2; :::; ik) = TM[i1
ik] where TM is a function which returns
the rightmost log2M bits of i1
ik. Since (
) is a boolean operation the values i1; i2; :::; ik must
be encoded in binary.
3. Error Correcting Codes (ECC) A declustering approach based on using error correcting codes was proposed
by Faloutsos et al . It works for binary attributes or an attribute where the number of partitions
on it, di, is a power of 2. For the binary case the problem is reduced to grouping the 2k binary strings on
k bits in M groups of dissimilar strings. The main idea is to form groups of strings such that each group
forms an Error Correcting Code (ECC). In case di is a power of 2, the binary representation of the domain
is used. Thus if each di can be represented as a binary string of lengthmthen we need to construct an ECC
on km bits out of which log2M bits will be parity check bits while the rest will be information bits.
4. Hilbert Curve Method (HCAM) A declustering method based on space filling curves was recently proposed
by Faloutsos and Bhagwat . Such a curve visits all points in a k -dimensional grid exactly once
and never crosses itself. Thus, it can be used to linearize the points of a grid. The authors use such a curve,
called theHilbertCurve  to fill the k -dimensional grid and then assign the disks to the buckets in a round
robin fashion. Thus, if H is the function which imposes the linear ordering generated by the Hilbert Curve
on the grid points (buckets) then diskOf (i1; i2; :::; ik) = H(< i1; i2; :::; ik >)modM.
2.3 Declustering and Optimality
Ideally, wewould like a declusteringmethod to be strictly optimal. Queries can vary frombeing completely specified,
partialmatch, range queries, correlational(diagonal) queries, etc. Consider a cartesian product file F which
has N;N > M total grid-blocks. Since the number of grid-blocks are greater than the number of disks, at least
two grid-blocks (buckets) will be mapped to the same disk. We can always come up with a query (using union)
which accesses exactly these two buckets. Thus, this particular query will not be optimal. This is independent
of the declusteringmethod used. Thus, it is not possible to have a declustering method which is strictly optimal.
2 0 1 3 1 3 2 0
3 1 0 2 0 2 3 1
0 2 3 1 3 1 0 2
1 3 2 0 2 0 1 3
3 1 0 2 0 2 3 1
2 0 1 3 1 3 2 0
1 3 2 0 2 0 1 3
0 2 3 1 3 1 0 2
1 2 1 2 1 2 1 2
0 3 0 3 0 3 0 3
3 2 1 0 3 2 1 0
0 1 2 3 0 1 2 3
3 0 3 2 1 0 3 0
2 1 0 1 2 3 2 1
1 2 3 2 1 0 1 2
0 3 0 1 2 3 0 3
0 1 2 3 0 1 2 3
1 2 3 0 1 2 3 0
2 3 0 1 2 3 0 1
3 0 1 2 3 0 1 2
0 1 2 3 0 1 2 3
1 2 3 0 1 2 3 0
2 3 0 1 2 3 0 1
3 0 1 2 3 0 1 2 3 2 1 0 3 2 1 0
2 3 0 1 2 3 0 1
1 0 3 2 1 0 3 2
0 1 2 3 0 1 2 3
0 1 2 3 0 1 2 3
1 0 3 2 1 0 3 2
2 3 0 1 2 3 0 1
3 2 1 0 3 2 1 0
Figure 1: A Declustering Example
From a practical viewpoint, however it can often suffice to consider range and partialmatch queries only since
these are the most commonly occurring class of queries in a database. Given that a declustering method cannot
be strictly optimal, it is thus desirable to have a declusteringmethod that is strictly optimal for partial match and
range queries. Much work has been done in proving results about performance bounds of partial match queries
for different declustering techniques    . Some results also exist about the conditions under which a
strictly optimal declustering can be achieved for partial match queries   .
Recent work  has derived sufficient and necessary conditions for optimality of a declustering technique
with respect to partial match queries when the number of partitions on all attributes are less than the number of
disks. The specific focus of  was on p -ary cartesian product files where (8i)(di = p). It was shown that
there is no strictly optimal allocation for a p -ary cartesian product file if
p(p2+p2)=2 M pd1 or p2 M pnp2p+2 1.
Thus, for all practical purposes the non existence of strictly optimal declustering methods with respect to
partial match queries was shown when the number of partitions on all attributes is less than the number of disks.
Since range queries are a superset of partial match queries these results hold for range queries too.
The above result while certainly of theoretical interest is not disheartening from a practical viewpoint. For
most medium to large databases having more partitions on an attribute than the number of disks is expected to
be quite common. For example, for a 16 disk system, with 8KB disk blocks, 64 bytes/record and 3 declustering
attributes in a relation, only 1 million records are needed in a relation to have 16 partitions on each dimension,
which is not unrealistic for a database requiring parallel processing.
The following discussion shows that range queries placemore constraints than partialmatch queries, and optimality
for them is harder to achieve. Specifically, a significant observation is that while the condition (8i)(di
M) guarantees optimality for partial match queries under many conditions, it does not do so for range queries.
Lemma 1. If M = ab is a composite integer and (9i; j)(di a + 1; dj b + 1) then a strictly optimal
declustering for range queries does not exist.
Proof: Refer .
Lemma 2. If M is a prime integer and (9i; j)(di 3; dj M) then a strictly optimal declustering for range
queries exists iff
(1)M = 1; 2; 3; 5 and d = 2 or
(2)M = 1; 2; 3 and d 3.
Proof: Refer .
Theorem 1. If d 3 then a strictly optimal declustering for range queries exists iff M = 1; 2; 3.
Proof: This is a direct consequence of Lemmas 1 and 2.
In the following table we summarize the main optimality results for various declustering methods.
Declustering Techniques and Optimality
Declustering Restriction on Restriction on Conditions on
Method Number of Disks Number of Partitions Optimal Queries
DM/CMD None None PM :Exactly one field
Range/PM: if one
of the range domains is an
integral multiple of M
FX power of 2 None PM: Exactly one field
Power of 2 PM: with an unspecified
attribute s.t. di M
ECC power of 2 power of 2 None derived
HCAM None None None derived
3 Latin Hypercube DeclusteringMethods (LHDM)
Latin Squares  are two-dimensional structures which show very good properties, and have been widely used
in experimental designs to ensure least redundancy and maximum coverage for the minimal experimental effort.
We generalize Latin Squares into higher dimensions and define a class of declustering methods called Latin Hypercube
Declustering Methods (LHDM).
Definition 9 [Latin Squares] A Latin Square of order n is an n n square composed of symbols from 0 to
n 1 such that no symbol appears more than once in a row or column . Zhou et al discuss some properties
of declustering methods using Latin squares in .
Definition 10 [Latin Hypercubes] ALatin Hypercube of dimension d and order n is an nn:::n hypercube
of dimension d composed of symbols from 0 to n 1 such that no symbol appears more than once in any row
for all dimensions.
Definition 11 [Latin Hypercube Declustering Methods (LHDM)] A declustering method which uses a Latin
Hypercube of dimension d and order M as its basic building block is called a Latin Hypercube Declustering
Method. The hypercubes are replicated along each dimension till they fill up the domain space of the relation.
In case the domain space in some attribute is not a multiple of M then the last hypercube in that dimension is
In the following discussion we use the term Latin Hypercube and Latin Hypercube Declustering Method interchangeably.
This is not to imply that the complete grid is mapped as a latin hypercube but that it is mapped
using a latin hypercube as a basic block. We now derive some basic properties of Latin Hypercubes and show
sufficient and necessary conditions for a method to belong to the class of Latin Hypercubes.
Definition 12 [Periodic Allocation] A declustering method is said to be periodic if
(8j 2 f1; 2; :::; dg)diskOf(< ii; i2; :::; ij; :::; id >) = diskOf(< ii; i2; :::; ij +M; :::; id >); ij +M dj
Definition 13 [Row Optimal Allocation] A declustering method is said to be row optimal if the declustering
method is optimal for all queries such that the length of the query is 1 on all but one declustering attribute.
Lemma 3. If a declustering method is row optimal, then it is periodic.
Proof: Refer .
Theorem 2 A declustering method belongs to the class LHDM iff it is row optimal.
Proof: Refer .
Corollary: DM/CMD, GDM, FX, and Latin Squares belong to class LHDM.
Proof: DM , GDM , FX , and Latin Squares , each have been shown to be row optimal in the
respective papers. Using Theorem 2, all these methods belong to the class LHDM. Q.E.D.
3.1 Performance Analysis of LHDM
In this section we analyze Latin Hypercube Declustering Methods and derive conditions under which optimal
parallelism is achieved. To help understand the performance of queries when these conditions do not hold, we
also derive upper bounds on the worst case behaviour of all queries. Finally, to understand the expected performance
of LHDM we analyze their average case behaviour on queries. All of these results are applicable to any
declusteringmethod which belongs to the classLHDM, e.g. CMD, FX,GDM, etc. The proofs to the the Lemmas
and Theorems in this section can be found in .
Definition 14 [Interval Domain Space] Any query on the cartesian product file F will have to access all the data
in the interval it intersects. Thus, the range on any dimension i, of a range query can be transformed to the coordinate
system of the interval domain 0 li di. We define this grid with the interval domains as its axes as
the Interval Domain Space.
Definition 15 [Hyper-rectangle] A Hyper-rectangleH is a subspace of the d-dimensional interval domain space
such that if intervals Iik; Iil intersectH on dimension i then 8(j)(k j l)Iij intersectsH. It can be observed
that any range query can be represented as a hyper-rectangle in the interval domain space.
Theorem 3. LHDM is query optimal for all range queries whose length on some dimension is equal tokM where
Note that Theorem 3 provides only sufficient conditions under which queries are optimal. Thus, it is possible
to have queries which do not satisfy this condition and are still optimal. Next we characterize a subset of such
Lemma 4. Let Q be a range query which needs to examine hyper-rectangle A = di
=1(Li; Li + li 1), where
0 Li di li and 1 li < M for 1 i d. Without loss of generality, let li1 li2 lid, where
lik 2 fl1; :::; ldg for 1 k d, Q is required to access at most Qd1
k=1 lik buckets on each disk.
Lemma 5. Let
A = di
=1(Li; Li + kiM + li 1);
A1 = (L1; L1 + k1M 1) (di
=2(Li; Li + kiM + li 1));
Al = (A Sl1
t=1 At)TRl for 2 l d
Rl = (l1
t=1(Lt; Lt + ktM + lt 1)) (Ll; Ll + klM 1) (dt
=l+1(Lt; Lt + ktM + lt 1));
Ad+1 = di
=1(Li + kiM; Li + kiM + li 1);
where 0 Li di kiM li, 0 < li <M for 1 i d.
A is a hyper-rectangle in S. Thus, all Ai’s are hyper-rectangles in S for 1 i d+ 1 and have the following
1. A = Sd+1
2. The length of Ai on dimension i is kiM for 1 i d.
3. Ai TAt = ; if i 6= t for 1 i; t d+ 1, where ; is empty set.
It is obvious that the hyper-rectangles in F required by any range query, which do not satisfy the condition of
Theorem 3, can be represented by A in Lemma 5. Theorem 4 characterizes a subset of such queries for which
LHDM is still optimal.
Theorem 4. Let Q be the same range query as A in Lemma 5. LHDM are optimal for Q if (1=B + lid=M) > 1,
where B = Qd1
j=1 lij and li1 li2 lid.
Theorem 5. For any range query Q required to examine P buckets, at most
dP=Me + (M 1)d1 1
buckets are accessed per disk in response to Q.
Assumption: For the following discussion we make the assumption (8i)(di = nM).
Lemma 6. Assume that for any attribute i, all ranges [Li; Ui); Li; Ui 2 Di, could occur with equal probability
in any range query. Now, the probability of any range query being optimal is at least
p = 1 nM2 (n 1)M 2
Clearly, we can make p large enough by properly selecting n. The probability of a range query not being optimal
is less than 1 p. The above result shows that the performance of LHDM improves with the dimensionality of
Let range query Q be required to examine hyper-rectangle A = di
=1(Li; Li + kiM + li 1) containing P
buckets, where 0 Li nM kiM li and 0 li < M. Assuming li’s are independently and uniformly
distributed in f0; 1; :::; M 1g, we have the following theorem.
Theorem 6. In response to the range query Q above, at most
dP=Me + (1 p) (M 1)d1
buckets are accessed per disk on the average.
Theorems 5 and 6 provide two upper bounds which provide insight into the expected behaviour of LHDM. However,
the bounds are not the tightest possible and hence the actual performance of LHDM can be much better.
Since theoretical analysis was rapidly getting intractable, we decided to carry out an experimental evaluation to
study the behaviour of LHDM in more detail. These are described in the next section. One of the most promising
applications of parallel databases is in decision support applications running against very large databases. In
such scenarios range queries are usually expected to examine a very big subspace of F, i.e. P in Theorems 5 and
6 will be very large. Thus dP=Me, the optimal number of disk accesses, is much greater than (M 1)d1 1
or (1 p)((M 1)d1=2d1 1). And hence, LHDM is expected to behave nearly optimally for most range
4 Experimental Evaluation
We believe that while theoretical studies such as     and that presented in the previous section,
provide valuable insight into the properties of declusteringmethods, the picture is not complete without a detailed
experimental evaluation. This is more so because of the fact that all declustering techniques are not amenable to
detailed theoretical analyses and the bounds obtained are not exact in most cases. Specifically, since in practice
no restrictions can be placed on the size and shape of queries, as well as the number of attributes or their domain
sizes, we believe an evaluation is needed which varies these dimensions as parameters and studies their effects on
the performance of various declustering methods. Thus, we have chosen to carry out experimental evaluations
to examine the performance of LHDM. The aim is to see how the different techniques belonging to this class
compare amongst themselves and also with other prominent techniques proposed in literature. We choose two
declustering methods from the class LHMD, namely FX and CMD, and two others namely ECC and HCAM,
for our experimental evaluation.
The main results of the experiments  are as follows: (i) various declustering methods proposed in literature
show a noticeable difference of performance (in relative terms though not much in absolute terms) for small
queries, (ii) for large queries, Latin Hypercubes perform very well, (iii) the performance of declusteringmethods
is quite sensitive to the query shape and LatinHypercubes show better performance for linear queries and (iv) the
deviation of most declustering methods from optimality decreases as the number of dimensions, i.e. the number
of attributes, of the query box is increased, and specially so for Latin Hypercubes.
Our overall conclusions are that (i) no declusteringmethod can be optimal for all queries on a large database,
(ii) for large databases and large queries Latin Hypercube methods perform very well and are not very far from
optimal, (iii) information about commonly posed queries can be useful in selecting onemethod over anothermuch
like physical database design in centralized and distributed databases, and this choice is crucial for small queries,
and (iv) since different methods may turn out to be better for different relations, based on the queries posed on
them, commercial DBMSs will have to support more than one declustering method, much like different kinds
of access methods and index structures in today’s databases. Based on our studies, future work in the area of
declustering must address issues such as (i) how do grid-based methods perform compared to non grid-based
ones, (ii) how do various methods perform when data skew and attribute correlation is present, and (iii) how can
information about query sets be used in selecting a declustering method appropriate for a relation.
5 Project Architecture, Status & Future Directions
In the last few years, three parallel database (software) architectures have been considered, namely shared-memory,
shared-disk, and shared-nothing . Shared-memory architectures suffer from scalability problems, and thus
grew out of favor. An initial consensus was that shared-nothing architectures are the most promising [2, 25],
though recently support has been expressed for the shared-disk architecture . We believe that as technology
is progressing, the distinction between the latter two is becoming blurred. This is because the shared-disk
parallel database architecture is most suited toMPPs, while shared-nothing architecture is most suited to NOWs.
However, sinceMPPs and NOWs are becoming comparable froma hardware viewpoint, i.e. aggregate processor
cycles, communication bandwidth and latency, and I/O bandwidth and latency1 , the distinction between sharednothing
and shared-disk database architectures is diminishing. Essentially, in any parallel architecture there are
going to be nodes with two kinds of capabilities, namely processing and I/O. Nodes that have both capabilities
would have separate processors for each. Nodes will be connected by means of a high bandwidth and low latency
network, whose specific topology will be largely irrelevant. We believe that software architectures for parallel
databases must keep in mind these trends in parallel hardware architecture.
PADMA has been an ongoing project for the last three years. We provide a brief overview of the project here.
A detailed description can be found in . Figure 2 shows the architecture of the PADMA parallel database
1Experience with I/O intensive applications has shown that CPU-controlled I/O is not a good idea, and a DMA or I/O processor is
certainly needed. Trends in architecture are similar, i.e. MPPs have dedicatedI/O processors,while nodesin NOWswith I/O capability
have DMAs. Coupled with the fact that NOWs will have high speed networks, accessing a remote node’s disk is going to be comparable
to that of accessing a disk connected to an I/O processor in a MPP.
Query Compilation, Optimization, Scheduling
Parallel Relational Algebra Layer;
Join, Union, Aggregate, Sort, Scan, etc.
Query, Data Retrieval
Parallel Record Management Layer
Concurrency Control, Buffer Management, Recovery
Simple Selection, Sort etc.
Query Optimization and Processing
. . .
. . .
. . .
Figure 2: Parallel Database Architecture
manager. The architecture is designed around the following hypotheses:
A geometric model, where points represent tuples, subspaces (boxes) represent queries, and geometric intersection
algorithms handle query processing, is a useful way to visualize relational database querying.
Given that datamovement, between disk andmemory, aswell as betweenmemories, is themain bottleneck,
effective data declustering is the key to performance.
Making the various layers of the database manager, e.g. record management, query processing and optimization,
etc., understand the declustering below can have significant payoffs for performance. Thus, the
DBMS layers must become declustering-aware.
Multi-threading of various DBMS functions is important for performance. Additionally, the mapping of
various threads to processorsmust be done in a declustering-awaremanner to take advantage of the affinity
certain processors may have for certain computation (e.g. due to data availability).
Oneway of viewing the results obtained in the PADMA project is as an ongoing experiment in testing the hypotheses
listed above. As of this reporting the experiment is not complete. One or more of the above hypotheses
have been tested to varying degrees. Others are part of our ongoing and future investigations.
The results  so far include (i) development of declustering techniques [14, 12] and their performance evaluation
, (ii) declustering-aware query processing algorithms , (iii) parallel database loading algorithms
, and (iv) parallel query optimization [23, 13]. We are currently building a main-memory prototype of the
Futurework in the PADMA project includes (i) detailed performance evaluation of various techniques developed,
(ii) extension of the parallel techniques developed for points to handle intervals and regions, for temporal
and spatial data, and (iii) development of example applications on top of the prototype .
PADMA represents the effort of various individuals over the last three years. We would like to acknowledge
the contributionsmade by Prof. Jian-Zhong Li of Heilongjiang University, P.R.C., Dr. Doron Rotemof Lawrence
Berkeley Laboratory, Sakuntala Kavuri of Intel Corporation, Gary Elsseser of the University of Minnesota, and
Sujal Parikh of CDAC, India. We would also like to thank the anonymous referees for their valuable comments.
This research has been supported in part by the National Science Foundation grant IRI-9110584. Technical reports
related to PADMA can be obtained by anonymous ftp from ftp.cs.umn.edu: /users/padma.
 H. Boral and et al. Prototyping Bubba: A highly parallel database system. IEEE TransactionsonKnowledge andData Engineering,
2(1), March 1990.
 D. J. DeWitt and J. Gray. Parallel database systems: The future of high performance database systems. Communications of the
ACM, 35(6):85–98, June 1992.
 D.J. Dewitt and et al. TheGamma databasemachine project. IEEE Transactionson Knowledge and Data Engineering, 2(1), March
 H.C. Du and J.S. Sobolewski. Disk allocation for cartesian product files on multiple disk systems. ACM Transactions on Database
Systems, pages 82–101,March 1982.
 C. Faloutsos and P. Bhagwat. Declustering using fractals. Parallel and Distributed Information Systems, pages 18–25, January
 C. Faloutsos and D. Metaxas. Disk allocation methods using error correcting codes. IEEE Transactions on Computers, pages
 O. Frieder. Multiprocessor algorithms for relational-database operations on hypercube systems. IEEE Computer, November 1990.
 T. Fujiwara, M. Ito, T. Kasami, M. Kataoka, and J. Okui. Performance analysis of disk allocation method using error correcting
code. IEEE Transactions on Information Theory, pages 379–384,March 1991.
 S.Ghanderharizadeh andD.J. DeWitt. A multiuser performance analysis of alternative declustering strategies. Proceedings ofData
Engineering Conference, Feb 1990.
 L. Harada,M. Nakano, M. Kitsuregawa, and M. Takagi. Query processing method for multi-attribute clustered relations. Proceedings
of International Conference on VLDB, pages 59–70, August 1990.
 B. Himatsingka and J. Srivastava. Performance evaluation of grid basedmulti-attribute record declustering methods. Proceedings
of 10th International Conference on Data Engineering, Feb 1994.
 B. Himatsingka, J. Srivastava, J. Li, and D. Rotem. Latin hypercubes: A class of multidimensional declustering techniques. Technical
Report TR 94-05, University of Minnesota, Minneapolis, Department of Computer Science, January 1994.
 B. Himatsingka, J. Srivastava, and ThomasM. Niccum. Tradeoffs in parallel query processing and its implications for query optimization.
Technical Report TR 94-09, University of Minnesota,Minneapolis, January 1994.
 Li Jianzhong, J. Srivastava, and D. Rotem. CMD: A multidimensional declusteringmethod for parallel database systems. Proceedings
of International Conference on VLDB, August 1992.
 A. S.Abdel-Ghaffar Khaled and Amr ElAbbadi. Optimal disk allocation for partial match queries. ACMTransactions on Database
Systems, pages 132–156,March 1993.
 M.H. Kim and S. Pramanik. Optimal file distribution for partial match queries. Proceedings of ACM SIGMOD, pages 173–182,
 J. Li, D. Rotem, and J. Srivastava. Algorithms for loading parallel grid files. Proceedings of ACM SIGMOD, May 1993.
 Thomas.M. Niccum, J. Srivastava and J. Li. Declustering Aware Parallel Join Algorithms. Proceedings of the International Conference
for Young Computer Scientists, Beijing, 1993.
 J. Nievergelt, H. Hinterberger, and K.C. Sevcik. The grid file: an adaptable, symmetric multikey file structure. ACM Transactions
on Database Systems, pages 38–71,March 1986.
 David A. Patterson. A quantitative case for networks of workstations (NOW). Cray Distinguished Lecture Series, April 1994.
 H. Pirahesh and et al. Parallelism in relational database systems: Architectural issues and design approaches. Proceedings of 2nd
International Symposium on Database in Parallel and Distributed Systems, Dublin, Ireland, July 1990.
 J. Srivastava. A parallel data management systemfor large-scale NASA datasets. Proceedings of the 3rd NASA GSFC Conference
on Mass Storage Systems and Technologies, pages 283–299, October 1993.
 J. Srivastava and G. Elsesser. Optimizing multi-join queries in parallel relational databases. Proceedings of the 2nd International
conference on Parallel and Distributed Information Systems, pages 84–92, January 1993.
 J. Srivastava, T. M. Niccum and J. Srivastava. PADMA: A PArallel DatabaseMAnager. Technical Report TR 94-47, University of
Minnesota,Minneapolis, Department of Computer Science, August 1994.
 M. Stonebraker. The case for shared nothing. Database Engineering, 9(1):4–9, 1986.
 P. Valduriez. Parallel database systems: The case for shared something. Proceedings of 9th International Conference on Data
Engineering, pages 460–465, 1993.
 Thorsten von Eicken, David E. Culler, Seth Copen Goldstein, and Klaus Erik Schauser. Active messages: a mechanism for integrated
communication and computation. Technical Report TR CSD-92-675, University of California, Berkeley, Computer Science
Division, June 1992.
 Yvonne Zhou, Shashi Shekhar, and Mark Coyle. Disk allocation methods for parallelizing grid files. Proceedings of 10th International
Conference on Data Engineering, Feb 1994.
Fault Tolerance Issues in Data Declustering for Parallel Database
Leana Golubchik Richard R. Muntz
UCLA Computer Science Department
Maintaining the integrity of data and its accessibility are crucial tasks in database systems. Although
each component in the storage hierarchy can be fairly reliable, a large collection of such components is
prone to failure; this is especially true of the secondary storage system which normally contains a large
number of magnetic disks. In designing a fault tolerant secondary storage system, one should keep in
mind that failures, although potentially devastating, are expected to occur fairly infrequently; hence, it
is important to provide reliability techniques that do not (significantly) hinder the system’s performance
during normal operation. Furthermore, it is desirable to maintain a reasonable level of performance under
failure aswell. Since high degrees of reliability are traditionallyachieved through the use of duplicate
components and redundant information, it is also reasonable to use these redundancies in improving the
system’s performance during normal operation. In this article we concentrate on techniques for improving
reliability of secondary storage systems as well as the resulting system performance during normal
operation and under failure.
Maintaining the integrity of data and its accessibility are crucial tasks in database systems. Consequently, the
reliability requirements of a database, and especially its storage hierarchy, are very stringent. A measure of a
storage system’s reliability is the mean time till it experiences loss of data due to the failure of one or more of
its components; because we are interested in a continuously operating system, we use the term “data loss” to
refer to inability to access data, due to failure, whether or not it is recoverable from archival storage and/or logs.
(Thus, in this article we do not discuss recovery of information through the use of full dumps, log files, etc.)
A database storage hierarchy typically contains a large number of disks, which not only provide the necessary
storage but also the bandwidth and/or the number of disk arms required to exhibit reasonable performance1. For
instance, transaction processing systems (i.e., with OLTP workload) are under stringent system responsiveness
requirements, e.g., 99 percent of all transactionsmust be completedwithin 1 second. Such systems are configured
according to the number of I/Os/second desired, rather than the number of MBs necessary to store the data .
Thus, the workload of the system greatly influences its storage configuration as well as (and we shall see this
later) the design of its reliability schemes.
Although a single disk can be fairly reliable, given a large number of disks, the probability that one of them
fails can be quite high. For example, if the mean time to failure (MTTF) of a single disk is 200; 000 hours, then
the MTTF of some disk in a 200-disk system is on the order of 1000 hours, i.e., a disk failure is expected (approximately)
once every 42 days. To improve the reliability and availability of the secondary storage system,
1A database storage hierarchy can also include a tertiary store; however, in this article, we limit our discussion to the reliability of a
two level storage hierarchy.
some form of data redundancy must be introduced. One way to introduce redundancy into the system is to use
parity based schemes  which construct a parity block for every d data blocks; the parity block plus the d data
blocks constitute a parity group. Whenever a disk fails, a data block on the failed disk can be reconstructed by
reading and computing the exclusive OR of the corresponding parity and data blocks. Examples of parity based
schemes include RAIDs , clustered RAIDs  and various parity striping schemes . Proper reliability
techniques can increase mean time to data loss (MTTDL) to millions of hours .
Full mirroring  is a special case of parity striping (with d = 1), where each disk is replicated on another
disk; whenever a disk fails, its mirror can be used to retrieve the missing data. The disk farm is composed of a
number of such pairs. Full mirroring has a higher storage overhead than other parity based schemes with d >
1 (e.g., RAID) because data is fully duplicated, but it can offer better performance in terms of throughput and
response time  than the parity based schemes. For instance, in , the authors exploit the availability of two
copies of the data to optimize seek times in a mirrored disks environment.
The amount of redundant information stored determines the storage overhead for providing reliability and the
system’s resiliency to disk failure. The storage overhead for parity based schemes is 1
d+1 of the total storage space.
In general, themore redundant information is stored, the lower is the probability that a failure results in data loss,
but the higher is the cost of providing reliability. Furthermore, the placement of the redundant information on the
disks [24, 25] influences the system’s behavior during normal operation and under failure as well as its ability to
recover quickly and return to the fully operational state. When designing a fault tolerance scheme, the following
aspects of the disk subsystemmust be examined: a) performance under normal operation (e.g., [7, 31]), b) mean
time to data loss (or systemfailure) (e.g. ), and c) performance of the disk subsystemunder failure, i.e., when
one or more disks are inoperable or inaccessible (e.g., [35, 30, 17, 19, 18]). We should keep in mind that failures
are expected to occur relatively infrequently, so most of the time a systemis in a fully operationalmode. Thus, it
is important to provide reliability techniques that do not (significantly) hinder the system’s performance during
normal operation. Since high degrees of reliability and availability are achieved through the use of redundant
information (and duplicate components), it is also reasonable to use these redundancies in improving the system’s
performance during normal operation, e.g., as in mirrored disk systems  (see Section 3 for more details).
In order to maintain a reasonable MTTDL, it is desirable to provide immediate repair of a failed disk; after
the first failure has occurred, there is a vulnerability window during which a second failure causes loss of data.
(We assume, as is true of all the schemes surveyed here, that for every disk failure, the additional failure of one
of the surviving disks can cause data loss.) To this purpose, “hot standby” disks (or spares) are often provided,
and the system is designed to automatically rebuild the contents of the failed disk on the standby disk, using the
redundant information on the surviving disks . The parity group size effects: a) the time required to rebuild
a failed disk (and therefore the MTTDL) and, b) the workload (measured in accesses per second per disk) that
can be supported during the rebuild process, and c) the system’s performance under failure. The MTTDL of a
RAID is easily shown to be inversely proportional to the rebuild time [6, 11]; in the RAID system described in
, rebuilding the failed disk contents at maximum speed (the capacity of the standby disk) results in the use
of the entire capacity of the surviving disks in the array. Thus rebuilding at maximum rate means that the array
can perform no other work during the rebuild period. One can of course, tradeoff the rebuild rate with the rate
at which the surviving disks process normal workload requests. However, this increases the time to rebuild the
failed disk contents and thereby decreases the MTTDL.
Although disk failures are infrequent, a single disk unavailability is still a relatively common occurrence as
compared to data loss (or system failure). Therefore the performance of the system under failure and especially
during the repair period (when the data on a failed disk is being rebuilt) is of concern. The RAID organization
achieves a low cost in redundant storage overhead, as compared to mirrored systems, but at the price of degraded
performance under failure 2. In the worst case (a workload of all reads and no writes) this can double the access
2Note that, RAID systems also pay a performance penalty during normal operation; this is due to having to write a parity block on
every write operation.
rate to the surviving disks and thus in effect, cut the capacity of the array in half. Consider for example a sharednothing
 database machine architecture, where each node contains one or more disk arrays. The impact of
a failure on the total system performance is dependent on the characteristics of the system workload; it is most
severe in the case of a “decision support” environment in which complex queries are common and the database
tables have been partitioned among the disks on all or many nodes, for the purpose of increasing I/O bandwidth.
Such complex operations are typically limited by any imbalance3 in the system, which can be caused either by a
skew in the workload  or by a disk array with a diminished capacity, due to a failure . For example, in a
one hundred disk system, a single failed disk represents a loss of only 1% of the raw I/O capacity of the system.
However, if the effect of the failure is a reduction in the capacity of the array (to which it belongs) by say 25%,
then this failure can cause a significant imbalance in the system, and the impact on aggregate systemperformance
can be considerable.
In this article, we discuss techniques for providing a high degree of reliability and availability in a database
system; these techniques can be divided into two basic categories, which are as follows: 1) full replication, which
includes schemes such as shadow disks, interleaved declustering, and chained declustering, and 2) parity based
redundancy, which includes schemes such as RAID (Redundant Arrays of Inexpensive Disks), clustered RAID,
and parity striped disk arrays. We also discuss the tradeoffs, associated with each of these techniques, with respect
to the following metrics: a) storage overhead (due to redundancy), b) mean time to data loss (MTTDL), c)
performance during normal operation, and d) performance under failure.
The remainder of the article is organized as follows. Section 2 points out the differences between physical and
logical replication. Section 3 discusses full replication schemes, and the advantages and disadvantages associated
with those, both in the context of reliability and performance. Section 4 presents a similar discussion, but in the
context of parity based schemes. Finally, Section 5 presents our concluding remarks.
2 Physical vs. Logical Redundancy
In general, data redundancy can be implemented on different levels within a database system. In particular, we
distinguish between (1) physical redundancy and (2) logical redundancy. In what follows, we discuss the differences
between physical and logical redundancy in the context of full replication schemes; however, similar
comments apply to parity based schemes, such as RAID systems.
With physical level replication the contents of one area of a disk aremirrored on an area of another disk (in the
classical mirrored disk system, one entire disk is mirrored by another entire disk). The I/O controller generally
handles the replication and higher levels of software, such as the query optimizer, are not concerned4, i.e., higher
levels of software just see a collection of reliable disks with some changes in performance characteristics. With
logical fragmentation as in the Teradata  and Gamma  shared nothing database machines, relations are
fragmented and relation fragments are stored on independent nodes of the system. Replication is visible to the
query processing software and ismanaged by the database systemsoftware. For instance, since read requests can
be serviced using either copy of the data, replication can be used for load balancing purposes (we elaborate on
this further in Section 3). For reads, load balancing decisions can be made by the query processing software, i.e.,
at the logical level5, as in , or they can be deferred until the time of the actual I/O operation, i.e., performed
by the disk controller at the physical level, as in .
Note that, the dynamic scheduling studies that are discussed in this article, specifically in the context of
chained declustering (see Section 3.3), can be applied to both physical and logical replication methods. There
3Such queries would typically be performed in a “fork-join” manner (on a shared-nothingmachine),where the performance is limited
by the “slowest” node participating in the computation.
4Similarly, in RAID systems, higher levels of software just see a disk, i.e., a (logical) reliable disk with some changes in performance.
5Note that, one can have logical level replication and not do dynamic load balancing, i.e., just use the replication for reliability and
(static) redistribution of load after failure.
are however significant problems associated with dynamic data sharing across multiple nodes of a system, e.g.,
concurrency control, and efficient use of buffer space [38, 37]. We do not address these problems here due to
lack of space. With respect to logical replication one can view such studies as an investigation of the potential
benefits of dynamic load balancing, particularly with respect to robustness to workload imbalance and disk failure.
Determining whether these benefits compensate for the overhead and complexity of logical level dynamic
scheduling is beyond the scope of this article. In the remainder of this article we will concentratemainly on physical
replication, with the exception of interleaved and chained declustering schemes6 discussed in Sections 3.2
and 3.3, respectively.
3 Full Replication
We first concentrate on systems that use full replication as a form of redundancy, and present three variations on
this idea: 1)mirroring or disk shadowing, 2) interleaved declustering, and 3) chained declustering. Since all three
schemes fully replicate the data, they differ only in the way the replicas are placed on the disks. This placement
affects both reliability and performance.
Disk shadowing [5, 2] refers to maintaining two (mirrored disk) or more (shadow set) identical disk images on
different disks,mainly for the purpose of providing a highly reliable disk subsystem. Aread request to the shadow
set can be satisfied by any disk in the set; a write request must be executed on each of the disks in the shadowset.
When a disk fails, the data is still available on the other disks in the shadow set. To replace the failed disk, the
data must be copied from one of the disks in a shadow set to a replacement disk. This can be done either offline
or online. Offline copying is fast, but requires losing availability of data during the copying process (this can be
on the order of minutes/GB). Online copying has the advantage of availability of data but can be much slower
than offline copying (on the order of several hours). During the copying process the disk subsystemis vulnerable
to a second failure; with only two disks in a shadow set, a second failure results in data loss. Furthermore, the
systemoperates at a degraded level of performance. This degradation in performance is due not only to the failure
of a disk, but also to the copying process, which results in an additional workload on the shadow set. The more
“aggressive” is the copying process, themore it interferes with the normal workload. However, the faster a failed
disk is replaced, the less likely we are to lose data and the shorter is this degraded mode of operation. Hence, it
is desirable to balance the speed of the copying process, with degradation of performance experienced by the
normal workload due to the copying.
There are several disadvantages to disk shadowing. Firstly, there is the cost. Mirroring has a 100% storage
overhead. This is not a severe problem if the expected workload is of the OLTP type. According to , OLTP
systems have stringent responsiveness requirements; in order to avoid long queues of requests for the data, the
disks in such systems are usually purchased for their arms and not for their capacity. Secondly, there is the “write”
overhead. Since a write request must be serviced by every disk in a shadow set, it is not complete until the last
disk has finished writing. Even if all the disks in a shadow set can start working on the request simultaneously,
the write request will still experience the largest value of seek-plus-latency of all the disks in the shadow set.
There are advantages to disk shadowing, besides high reliability, which should be considered when comparing
its cost to the cost of parity based schemes. One such advantage, perhaps not an obvious one, is performance.
Withmultiple data paths, a shadow set can service several read requests in parallel, thus improving the throughput
of the disk subsystem. Furthermore, expected seek times for read requests can be improved by chosing the disk in
the shadow set with theminimum seek distance [5, 4]. This leads to a need for disk scheduling policies to exploit
these possibilities. Such policies for mirrored disk subsystems are studied in ; disk scheduling policies for
6These schemes were originally suggested as logical level schemes; thus we discuss them in that context.
real-time applications using mirrored disks are studied in . One interesting question that is addressed in  is
whether it makes sense to have more than 2 disks in a shadow set. The authors argue that two copies are sufficient
to provide a high degree of reliability, but that more than two copies can result in significant performance
3.2 Interleaved Declustering
In [3, 11] interleaved declustering is considered as a replication scheme at the logical level (see Section 2). It
can also provide an alternative to themirroring7 scheme, if applied at the physical level. We briefly describe this
scheme, which is illustrated in Figure 1, applied to physical level replication. The secondary storage subsystem
Figure 1: Interleaved Declustering
is divided into disk clusters, each of size N, e.g., in Figure 1, N = 4. Each file or table,R, is allocated equally to
each cluster; then each part assigned to a cluster is divided into N fragments. At all times two copies of this file
or table exist, termed primary copy and backup copy; both copies reside on the same cluster. The primary copy
of each fragment resides on one of the disks in a cluster, and the backup copy of the same fragment is divided
equally among the remaining N 1 disks of the cluster. During normal operation, read requests are directed to
the primary copy8 and write requests are directed to both copies (as in the mirrored disks case). When a failure
occurs, for instance of disk 1 in Figure 1, the read workload that was destined for disk 1 can be distributed among
the survivingN 1 disks of the cluster in which the failure occurred. This is an improvement over the mirrored
disks scheme where the additional workload, that was destined for the failed disk, ends up on a single surviving
disks (i.e., mirroring is a special case of interleaved declustering with N = 2).
Thus, interleaved declustering has the same storage overhead as mirroring, but it offers better performance
degradation properties, when a single disk failure occurs. The larger the cluster size, the smaller is the imbalance
in the workload (in the event of failure) between the fully operational clusters and the cluster with a failure9.
However, as the cluster size increases, so does the probability of two failures in the same cluster. Two failures in
any one cluster render data unavailable. Hence, the use of mirrored disks offers a higher level of reliability than
interleaved declustering (i.e., schemes withN >2)10.
7In the remainder of Section 3 we make comparisons to the mirrored disk scheme (i.e., shadowsets with 2 disks only), since it incurs
the same storage overhead as interleaved declustering and chained declustering (discussed in Section 3.3).
8Note that, it is possible to use both copies of the data to service read requests; however, with logical level replication, concurrency
control and buffer management issues must be considered (see Section 2).
9This could be a significant problem, for instance, in a shared-nothing databasemachine (see Section 1), such as the DBC/1012 ,
where the performance of the “slowest” node limits the performance of the entire system.
10Note that, this argument is an approximation, i.e., it only takes into consideration combinations of 2 failures. To make precise calculations,
we must take into consideration combinations of 3 or more failures; however, these are much less probable than combinations
of 2 failures.
3.3 Chained Declustering
In , chained declustering is considered as a replication scheme at the logical level of a shared nothing database
machine. This scheme can also provide an alternative to the classical mirroring scheme when applied to physical
level replication, as well as to the interleaved declustering scheme described in [3, 11]. We briefly describe the
concept of chained declustering from .
Chained declustering has the same storage overhead as compared to the classic mirroring scheme and interleaved
declustering, but, like interleaved declustering, it offers better performance degradation properties when
a single disk failure occurs. Figure 2 illustrates the chained declustering concept. Assume a file R is declustered
disk 0 disk 1 disk 2 disk 3 disk 4 disk 5 disk 6 disk 7
R0 R1 R2 R3 R4 R5 R6 R7
r0 r1 r2 r3 r7 r4 r5 r6
Figure 2: Chained Declustering
into M fragments, where M is the size of a disk cluster (e.g., in Figure 2, M = 8). At any point in time, two
physical copies of this file, termed the primary copy and the backup copy, are maintained. If the primary copy
of a fragment resides on disk i, then the backup copy of that fragment resides on disk i+1 (mod M). During
the normal mode of operation, read requests are directed to the primary copy11 and write operations update both
copies. When a disk failure occurs (e.g. disk 1 in Figure 2), the chained declustering scheme is able to adjust
the additional read workload to both copies of the data in such a way as to balance it evenly among the surviving
disks; this results in a less degraded performance (see  for more details).
There are several ways to perform the load adjustment depending on table declustering methods, storage organization,
and access plans. Since data is logically replicated, the query scheduler chooses an access plan in
order to balance the load. This form of load balancing has several limitations: (1) the load is only approximately
divided among the nodes; the assumption that a uniformdivision of the data corresponds to a uniformdivision of
the load can be incorrect with skewed reference patterns and (2) both short term and long term reference patterns
change with time and a static balancing scheme can not adjust to variations in load. Another way to balance the
load of the system is to apply some dynamic load balancing scheme, since it can adjust the load on each node
in real time to respond to statistical variations12. As already mentioned, several dynamic balancing schemes are
discussed in , in the context of mirrored disks systems. In , authors investigate the degree to which a dynamic
load balancing disk scheduling algorithm in conjunctionwith chained declustering can respond robustly to
variations in workload and disk failures (which destroy the symmetry of the system and introduce skewed load);
they demonstrate that simple dynamic scheduling algorithms can greatly improve the average response time as
compared to static load balancing.
Chained declustering has the same storage overhead as mirroring and interleaved declustering. But, it has
a higher reliability than interleaved declustering (but not as high as mirroring) . In order to lose data in the
chained declustering scheme (refer to Figure 2), two consecutive disks in the same cluster must fail. Note that
the probability of two consecutive disks failing in the same cluster, forM >2, is independent of the size of the
cluster. Hence, in the case of chained declustering, constructing a single cluster out of all the disks in the system
does not hinder the system’s reliability, but it can offer better load balancing in the event of failure. Since there is
no reliability penalty for using large clusters, the increase in load, due to a failure, can bemade as small as desired
by increasing the cluster size. This is not the case for interleaved declustering (as already mentioned in Section
11As with interleaved declustering, it is possible to use both copies of the data to service read requests; however, with logical level
replication, concurrency control and buffer management issues must be considered (see Section 2).
12Dynamic load balancingwould result in additional complexity in query processing software (see Section 2), e.g., in terms of concurrency
control; such complexity can be expensive, and consequently, dynamic load balancing schemes might be more suitable for large
queries, such as found in decision support type workloads, rather than OLTP type workloads.
3.2). Thus, chained declustering (forM > 2) offers better load balancing than either mirroring or interleaved
declustering, since it is able to distributed the additional load (due to failure) among all the disks in the storage
subsystem as opposed to a single disk (as in the case of mirroring) or the disks in a single cluster (as in the case
of interleaved declustering)13.
4 Parity Based Schemes
As already mentioned in Section 3, full replication schemes have the disadvantage of a 100% storage overhead.
To remedy this problem, we can use a parity based scheme. In this section we discuss three variations on such
schemes: 1) redundant array of inexpensive (or independent) disks (RAID) , 2) clustered RAID , and 3)
parity striping . We should note that there exists another variation on the RAID idea, termed RADD (Redundant
Array of Distributed Disks), which is a distributed version of a RAID5 system (refer to Section 4.2 for
a discussion on RAID5); we do not discuss it here due to lack of space but refer the interested reader to .
4.1 Disk Array Basics
The basic organization of an N +1 disk array is illustrated in Figure 3, where there is a cluster of N +1 devices
with N data devices and one parity device (N = 3). A file R is fragmented into blocks of size s, termed the
disk 1 disk 2 disk 3
d xor 5’
Figure 3: Basic RAID Organization
interleave unit size or the stripe unit, which is the amount of logically contiguous data that is placed on a single
device, e.g., d0 in Figure 3. The file is then interleaved among the N data devices, where N is the stripe width.
Each set of N data blocks is protected by one parity block; for instance in Figure 3, p0 = d0 d1 d2.
In general, there are three modes of operation for a disk array : 1) normal mode, where all disks are
operational, 2) degraded mode, where one (or more) disks have failed, and 3) rebuild mode, where the disks are
still down, but the process of reconstructing the missing information on spare disks is in progress. Under normal
operation read requests are directed to the disks holding the appropriate data14. A “small” read operation would
result in a single disk access, and a “large” read operation would result in a full stripe access (i.e., involving all
the disks in the cluster, except the parity disk). Every write request involves an access to at least 2 disks, due to
the necessary parity update. For instance, to replace d5 by d
5 in Figure 3, we must read d5 and p1, from disks 3
and 4, respectively, then compute the new parity, p
1 = d5p1 d
5, and then write out d
5 and p
1, to disks 3 and
4, respectively. Hence, a “small” write operation, involving data on a single disk, results in 4 I/O accesses, two
13We should note, that although there is no reliability penalty associated with using large clusters in a chained declustering scheme,
there are potential performance penalties. For instance, if the size of the fragments becomesvery small (whichwould happenif a relatively
small file was distributed over many disks), then some types of queries would have to be serviced by accessingmultiple disks, and this
can result in increased overhead .
14This could involve one or more disks of the array, depending the granularity of the stripe unit and the size of the request .
reads and two writes. A “large” write operation would be a full stripe access and result in a write on every disk
in the cluster, i.e., there is no need to read the old parity or the old data (e.g., p
0 = d
After a failure occurs the system continues to operate but in a degraded mode. For instance, suppose disk
3 fails in the system of Figure 3; then, to service a read request destined for the failed disk (e.g., d2), we must
read a full stripe in order to reconstruct the missing data, (e.g., d2 = d0 d1 p0). To service a write request
destined for the failed disk, we must do one of the following things. If the write request is for a data block, then
we must read the full stripe, to reconstruct the missing block, compute the new parity, and write the new parity.
If the write request is for a parity block, then it can be ignored. These additional full stripe reads and writes that
are necessary to reconstruct the missing data, result in a degraded performance of the disk subsystem. Note, that
the above description of servicing reads and writes destined for the failed disk is relevant to “small” reads and
writes only. “Large” read and write requests are full15 stripe operations regardless of whether there is a failure
or not. (This ignores edge effects of “large” operations, i.e., 11
2 stripes, for example.)
To reconstruct the missing data, i.e., enter the rebuild mode, we need a spare disk. Having a hot spare, i.e., a
spare disk that is online and ready for reconstruction as soon as a failure occurs, would significantly decrease the
vulnerability period, i.e., the period in which another failure would result in loss of data; decreasing this period
is also desirable because of the degraded system performance under failure. The basic reconstruction procedure
works as follows. A full stripe is read from all the surviving disks in the cluster, including the parity block. Then
themissing data block (fromthat stripe) is computed andwritten out to the spare disk. For instance, to reconstruct
d2 in Figure 3, we read d0; d1, and p0, and then compute the missing data, d2 = d0 d1 p0. Finally, d2 must
be written out to the spare disk.
Before discussing disk arrays in more detail, we present a list of design issues which should be considered
when constructing parity devices (in the following sections we address some of these issues in more detail):
a)redundancy support for hardware in addition to redundant information, e.g., multiple controllers, b) independence
of device failure is important since I/O subsystems require support hardware that is shared among multiple
disks (see Section 4.6), c) array size (or cluster size) affects the reliability of the system as well as its performance
in the normal and degraded modes of operation (see Sections 3 and 4.4), d) stripe width (or parity group
size) in the traditional RAID organization (see Section 4.2) is equal to the cluster size, whereas in Section 4.4
we show how the system’s performance under failure can be improved by relaxing this condition, e) interleave
unit size (stripe granularity) determines the number of devices that are involved in an access, and and hence it
affects the system’s performance during normal operation (we do not discuss this any further due to lack of space
but refer the interested reader to  for a performance comparison between byte interleaved, block interleaved,
and mirrored systems under normal operation), f) number of spares affects the reliability of the I/O subsystem
(see Section 4.3), and g) reconstruction time (or vulnerability window) is of crucial importance, because a system
operating under failure is not only vulnerable to a second failure (which results in a system failure, i.e., loss
of data) but it also exhibits degradation in performance; to reduce the MTTF of the whole system, it is necessary
to rebuild the failed disk as soon as possible but without significantly slowing down the normal workload;
in other words, the availability of data after a failure would not mean much if this data can not be accessed in a
“reasonable” amount of time (see Section 4.5 for a discussion of several reconstruction schemes).
4.2 RAID Organizations
In this section we describe the different RAID organizations, as they are presented in . Firstly, we present
the terminology16: 1) RAID1: is a data mirroring scheme, i.e., it uses full replication (see Section 3), 2) RAID2
15A large read doesn’t involve an access of the parity disk, under normal operation. The failure’s affect on system’s performance
depends on the RAID organization used; e.g., there would be no impact on the performance of a RAID3, because it uses the rotationally
synchronized byte interleaved organization which does not allow multiple parallel accesses anyway (see Section 4.2).
16We do not describe the RAID1 scheme in more detail, since it is very similar to the full redundancy schemes discussed in Section
3. The RAID2 organization uses Hamming code as its ECC, where some fraction of the redundant information is used to detect which
& RAID3: are parity based, parallel access schemes, where all the disks in a cluster are rotationally synchronized,
and 3) RAID4 & RAID5: are parity based, independent access schemes, where all the disks in a cluster
can simultaneously perform independent accesses. The synchronized RAID3 organization is traditionally byte
interleaved, as in [22, 31]. This is due to the common assumption that rotationally synchronized disks do not perform
independent accesses; hence, they are viewed as a single unit, withN rate of a single disk, and which can
satisfy one request at a time. (An exception to this view is the work presented in , where the authors describe
workloads under which it would be beneficial to use larger striping units in synchronized, i.e., RAID3, disk array
organizations.) The advantages of a traditional byte interleaved RAID3 are: 1) high bandwidth, 2) high reliability,
and 3) its performance in degraded modes (since every request results in a full stripe access, its performance
in degraded mode is equivalent to its performance in normal mode). A disadvantage of RAID3 is that it has low
throughput on small requests, since every request involves all the disks in cluster, no matter how large or small.
To remedy the problem of low throughput on small accesses, we can use the RAID4 and RAID5 schemes,
which both use block17 interleaving and can independently service multiple requests in parallel. The difference
between the two schemes is in the parity placement18. In the RAID4 scheme there is a dedicated parity disk,
as in the example of Figure 3. The problem with this arrangement is that the parity disk can become a bottleneck,
since every small write operation requires the reading and writing of parity. To remedy this problem, the
RAID5 scheme rotates the parity among all the disks in a cluster; this is illustrated in Figure 4. The basic idea
disk 1 disk 2 disk 3 disk 4
Figure 4: RAID5 Organization
is that RAID4 and RAID5 should still provide the high access rate of RAID3 on large requests but are also able
to provide high throughput on small requests. However, we should note that RAID4 and RAID5 suffer from
performance degradation on “small” write requests, since each non-full stripe write request results in four I/O
operations; due to lack of space, we do not discuss this problem here but refer the interested reader to [27, 26, 33].
As mentioned earlier, reconstructing a failed disk as soon as possible contributes significantly to improving the
MTTDL. Of course, to reconstruct a disk, we need a spare one. If the spare disk is offline, i.e., requires human
intervention, then the time to order it, install it, etc. will likely dominate the actual reconstruction process. However,
if the spare disk is online (i.e., a hot spare), then the vulnerability period19 of the system is determined by
the efficiency of the reconstruction process. The various approaches to improving the reconstruction process are
discussed in Section 4.5. In this section, we first address the question of “how many spares do we need?”. In
, the authors address this issue by simulating a disk array with 7 parity groups (or strings as they are called
in ) and varying the size of the spare disk pool. The basic result is that (with or without hot spares) there is
disk has failed (only one parity disk per cluster is necessary to correct the failure). Since most disk controllers can detect which disk has
failed, this is not necessary. Thus, we do not discuss the RAID2 organization any further.
17What is the desirable block size depends on the system’s expected workload (e.g., see ).
18Performance consequences of several parity placement schemes for RAID systems are investigated in [24, 25], where the authors
show that, for certain types of workloads, a “proper” choice of parity placement can result in a significant performance improvement.
19The period during which another failure results in data loss.
essentially no difference (with respect toMTTDL) between a spare pool of 7 disks and an “infinitely” large spare
pool, i.e., it is sufficient to provide one spare disk per parity group.
Another way to use spare disks to improve system performance, both during normal operation and under
failure, is to use a distributed spare  (instead of a dedicated spare). The basic idea is to use the spare disk
under normal operation to construct anN +2 (instead of anN+1 array) with spare blocks on all N +2 disks; an
example of a systemusing a distributed spare is illustrated in Figure 5. Advantages of an array with a distributed
disk 1 disk 2 disk 3 disk 4
Figure 5: Distributed Spare (N+2 Array)
spare are as follows: a) better performance under normal operation, since we are able to use N + 2 instead of
N + 1 disks, b) better degraded mode performance, since we are able to use N + 1 instead of N disks plus less
data is lost due to failure (since the spare blocks had no data), c) shorter reconstruction process (since less data
is lost due to failure), and d) higher probability that the spare is operational when it is needed, since it is being
used during normal operation (see Section 4.6 for discussion on infant mortality of disks). The disadvantages
are: a) when a new disk becomes available (to replace the failed one), there is a need for a “copy back” process,
i.e., copying of data to the new disk in order to create a distributed spare again, which could be done when the
systemis idle, and b)withN+2 disks in an array, there is a greater probability of a single disk failure, and hence,
distributed spare systems tend to spend more time in degraded performance modes.
4.4 Clustered RAID
It is desirable for the system to spend as little time as possible in the degraded mode of operation, because during
that period: a) the system is vulnerable to a second failure, which can result in data loss and b) the system
performance is degraded due to the failure. One way to improve the system’s performance under failure and at
the same time speed up the reconstruction process is to use the clustered array organization, proposed in .
The basic idea behind clustered disk arrays is to relax the assumption that the group size, G, should be equal to
the cluster size, C, where “group” refers to the parity groups size, i.e., the number of data blocks plus the parity
block, and the “cluster” size refers to the number of disks over which the parity group blocks are distributed. In a
traditional RAID architecture, as in , it is assumed that the group size is always equal to the cluster size. An
example of a system where the group size (G = 4) is less20 than the cluster size (C = 5) is illustrated in Figure
6. To place each parity group (i.e., three data blocks plus one parity block) on the disks, we must select 4 out of
5 disks in the system. Since there are (5
) ways to make such a selection in Figure 6, there are five possible types
of parity groups21.
The clustered organization does not require additional disks, since the overhead for storing redundant information
is determined by the group size,G; furthermore,Gdetermines the number of reads thatmust be performed
to reconstruct a data block from a failed disk. On the other hand, the MTTDL is determined by the cluster size,
C, since any two failures in one cluster result in data loss. Note that, there are benefits in choosing a cluster size
20Of course, the group size has to be at most as large as the cluster size, otherwise, the array would not able to recover even from a
21Note that in Figure 6, each group type appears to have a column of empty blocks; this is done for ease of illustration, i.e., the figure
illustrates the logical organization of the data on the disks rather than the physical one.
Group type 2 Group type 1
disk 2 disk 3 disk 4 disk 5
Figure 6: Clustred RAID Organization
that is greater than the corresponding group size, and they are as follows. When a disk fails,G1 blocksmust be
read, from C 1 surviving disks in order to reconstruct each block of the missing data. By properly distributing
the groups among all the disks in the cluster, the additional load, due to failure, can be distributed evenly over
all C 1 surviving disks. If r is the fraction of accesses in the normal workload that are reads, then the increase
in the workload due to one failed disk is determined by rG1
C1 . Hence, an array withG < C would perform better
under failure and would have a shorter reconstruction process. An analysis of clustered array’s performance
under failure, using three different reconstruction schemes (see Section 4.5) can be found in ; this analysis
indicates that there are significant advantages to using the clustered disk array scheme. There remains one problem
with respect to implementing the clustered array architecture, which is left open in . This is the problem
of computing, for a given data block, the location of its “buddy” data blocks and parity block (i.e., the rest of the
blocks in the parity group), which is addressed in [17, 29].
4.5 Recovery Procedures
Several reconstruction schemes are suggested in ; these include: a) basic rebuild, where the data is read from
the surviving disks, reconstructed through a parity computation, and thenwritten to the spare disk, b) rebuild with
read-redirect, where, in addition, read requests, for the portion of the data on the missing disk that has already
been reconstructed on a spare, are redirected to the spare disk, and c) piggy-backing rebuild, which takes advantage
of read requests for data on surviving disks and uses the retrieved information to reconstruct some portion
of the failed disk. In all three schemes, the authors  suggest that the write requests to the failed disk should
always be redirected to the standby disk. In  the authors question this decision and investigate another recovery
algorithm, in addition to the three proposed in , which they refer to as the minimal-update algorithm;
in this scheme, updates to the failed disk are ignored, whenever possible. A simulation of all four reconstruction
algorithms reveals that the two more complex schemes, i.e., read-redirect and piggy-backing, do not consistently
reduce the length of the reconstruction period. In particular, in light to moderate loads with G1
C1 < 0:5, the
schemes with no redirection result in a shorter reconstruction period. The reason  is that the benefits of offloading
the surviving disks do not outweigh the penalty of loading the replacement disk with random workload,
unless the surviving disks are highly utilized.
Several other issues should be considered when designing a reconstruction process, for instance, the size of
the reconstruction unit, which can be a track, a sector, a cylinder, etc. The tradeoffs are as follows. A larger reconstruction
unit should speed up the reconstruction process, however, it should also result in greater degradation
of performance, as experienced by the normal workload, i.e., the longer it takes to read a reconstruction unit, the
(possibly) greater is the queueing delay experienced by the normal workload. Another way to reduce the reconstruction
period is to start multiple (independent) reconstruction processes in parallel. In , the authors note
that a single reconstruction process (or in lock step reconstruction)22 is not always able to highly utilize a disk
array, especially when G1
C1 is relative small; in that paper, the authors investigate the benefits of using an 8-way
parallel reconstruction process23.
4.6 Independence of Disk Failures
Until now, we have primarily considered the failure of disks. However, there are other components in the I/O
subsystem that deserve attention, such as controllers, power supplies, cabling, cooling systems, etc. In , the
authors point out that such support hardware is normally shared by a disk string (all the disks on one bus), as
illustrated in Figure 7(a). A failure of one such shared hardware component, e.g., a power supply, would result
(a) Orthogonal Parity Group Organization
on disk 1
on disk 2
(b) Disk Matrix
Bus 1 Bus 2 Bus 1 Bus 2 Bus 3 Bus 4 Bus 5 Bus 3 Bus 4 Bus 5
Figure 7: Independence of Disk Failure
in the inaccessibility of an entire string of disks. Thus, disks sharing the same support hardware should not belong
to the same disk array. In fact, the disk arrays should be constructed orthogonally to the support hardware groups
. In , the authors compare theMTTDL of an array with an non-orthogonal organization to that of an array
with an orthogonal organization and show a significant improvement in reliability.
In addition to guarding against multiple failures due to a single support hardware failure, we would also like
to have an even load distribution, over all the disks in the system, when a failure does occur. However, the orthogonal
organization described above does not exhibit this property. Note that in that organization, a disk failure
creates an addition load only on the disks belonging to the same disk array as the failed disk. In , the authors
propose another approach, termed a disk matrix24, which also guards against single points of failure but with an
additional benefit of evenly distributing the additional load due to a failure over all the disks in the system. In
general, all blocks belonging to the same parity group (i.e., data blocks plus a parity block) are distributed among
the disks in the disk matrix according to the following rules: 1) no two blocks from the same parity group end
up on the same disk string and 2) the increase in the load due to a disk failure is evenly distributed among all the
disks in the matrix ; this is illustrated in Figure 7(b). Due to a lack of space we do not describe this scheme
any further but refer the interested reader to . We do note, however, that one disadvantage of this scheme, as
compared to the orthogonal organization scheme, is that it has a lower reliability, since essentially, it uses larger
22By a single reconstruction process (or in lock step reconstruction) we mean a recovery procedure where the reconstruction of one
data block must be completed before the reconstruction of another data block can begin.
23The parallel reconstruction process requires additional buffer space to hold the data blocks that have been read from the surviving
disk, but have not (yet) been used to reconstruct the missing data.
24The disk matrix is a generalization of the clustered disk array idea.
4.7 Parity Striping
In , the authors point out why traditional RAID5 organization  might not be the best solution for all types
of workloads, and more specifically for OLTP workloads (i.e., workloads with relatively small accesses). The
reason is that OLTP systems can not afford to use several disk arms on a single transfer, because the reduction
in (an already fairly short) transfer time can not offset the overhead associated with parallel transfer, such as an
increase in seek plus latency time (due to using multiple arms). Therefore, the authors propose another striping
scheme, termed parity striping, which can provide cheap reliable storage and high throughput. The basic idea
behind parity striping is to make a N + 2 disk array look like N + 1 logical disks plus a spare disk, rather than
as one logical disk (as in a RAID architecture). To this end, only parity blocks (rather than files) are striped
across all the disks in the system. Such a system is illustrated in Figure 8, where, for instance, blocks p20 and
p21 represent one contiguous parity segment which holds parity information for data blocks d00; d01; d10, and
d11 (where blocks d0i belong to file 0 and blocks d1i belong to file 1). Thus, a parity striping architecture allows
disk 1 disk 2 disk 3
d00 d10 d20
d21 d01 d11
Figure 8: Parity Striping
each small (relative to the size of the parity segment) access to be satisfied by a single disk, but it still provides the
reliability of a RAID5 system. In , a systemusing parity striping is analyzed and its performance is compared
to that of a system using mirrored disks and a system using RAID5. Another comparison of RAID5 and parity
striping performance (under normal operation) can be found in . Due to a lack of space, we do not discuss
these works here.
In summary, we have discussed two basic categories of schemes that store redundant information for the purpose
of reliability; these are: 1) full replication schemes and 2) schemes using parity information. In general full replication
schemes exhibit higher reliability and better throughput under normal operation (if both copies of the data
are used to service read requests). On other other hand, schemes using parity information have a much lower
storage overhead. The reliability characteristics of (most) schemes presented in this article can be summarized
briefly, as follows. To lose data in a system with D disks, the following must happen: 1) with mirroring two
disks must fail in the same mirrored pair, and there are D
2 such combinations, 2) with interleaved declustering
two disks must fail in the same cluster, and there are (C
C such combinations, where C is the size of each
cluster, 3) with chained declustering two consecutive disks in the same cluster must fail, and there are C = D
such combinations, where C is the size of the cluster (recall, that in chained declustering there is no reliability
penalty due to larger clusters, and in addition, there is a benefit to having larger clusters, namely the reduction
in additional load due to failure. Thus, it is (usually) desirable to have all the disks belong to the same cluster;
hence, C = D above), 4) with traditional RAID two disks must fail in the same cluster of size C = G (where
G is the parity group size), and there are (G
G such combinations, 5) with clustered RAID two disks must fail
in the same cluster of size C (where G C is the parity group size), and there are (C
 Fault Tolerant Disk Drive Matrix, Patent 5,303,244, Granted April 12, 1994. AT&T Global Information Solutions.
 NonStop SQL, A Distributed, High-performance, High-reliablity Implementaion of SQL. Technical Report No.
82317, Tandem Database Group, March,1987.
 DBC/1012 database computer system manual release 2.0. Technical Report Document No. C10-0001-02, Teradata
Corporation, Nov 1985.
 D. Bitton. Arm scheduling in shadowed disks. COMPCON, pages 132–136, Spring 1989.
 D. Bitton and J. Gray. Disk shadowing. VLDB, pages 331–338, 1988.
 P. Chen. An evaluation of redundant arrays of disks using an Amdahl 5890. Technical Report UCB/CSD 89/506, UC
Berkeley, May 1989.
 P. Chen, G. A. Gibson, R. H. Katz, and D. A. Patterson. An evaluation of redundant arrays of disks using an Amdahl
5890. ACM SIGMETRICS Conference, pages 74–85, 1990.
 Peter M. Chen and David A. Patterson. Maximizing Performance in a Striped Disk Array. ISCA, pages 322–331,
 S. Chen and D. Towsley. Performance of a mirrored disk in a real-time transaction system. ACM Sigmetrics 1991,
pages 198–207, 1991.
 S. Chen and D. Towsley. The Design and Evaluation of RAID5 and Parity StripingDisk Array Architecture. Journal
of Parallel and Distributed Computing, pages 58–74, 1993.
 G. Copeland and T. Keller. A Comparison of High-AvailabilityMedia Recovery Techniques. ACM SIGMOD Conference,
pages 98–109, 1989.
 David J. Dewitt, R. Gerber, G. Graefe, M. Heytens, K.Kumar, and M.Muralikrishna. Gamma : A high performance
dataflow database machine. VLDB Conference, pages 228–240, 1986.
 S. Ghandeharizadeh and D. J. DeWitt. Hybrid-Range Partitioning Strategy: A New Declustering Strategy for Multiprocessor
Database Machines. VLDB, pages 481–492, 1990.
 Garth A. Gibson. Performance and Reliability in Redundant Arrays of Inexpensive Disks. 1989 Computer Measurement
Group (CMG) Annual Conference Proceedings, December 1989.
 Leana Golubchik, John C.S. Lui, and Richard R. Muntz. Chained declustering: Load balancing and robustness to
skew and failure. RIDE-TQP Workshop, February 1992.
 Jim Gray, Bob Horst, and Mark Walker. Parity striping of disk arrays: Low-cost reliable storage with acceptable
throughput. VLDB Conference, pages 148–172, 1990.
 M. Holland and G. A. Gibson. Parity Declustring for Continuous Operation in Redundant Disk Arrays. In 5th Conference
on Architectural Support for Programming Languages and Operating Systems, Boston, MA, October 1992.
 M. Holland, G. A. Gibson, and D. P. Siewiorek. Architectures and Algorithms for On-Line Failure Recovery in Redundant
Disk Arrays. Submitted to the Journal of Distributed and Parallel Databases.
 M. Holland, G. A. Gibson, and D. P. Siewiorek. Fast, On-Line Failure Recovery in Redundant Disk Arrays. In 23rd
Annual International Symposium on Fault-Tolerant Computing, 1993.
 H. Hsiao and D. J. DeWitt. Chained Declustering: A New Availability Strategy for Multiprocessor Database Machines.
Proc. of Data Engineering, pages 456–465, 1990.
 R. Katz, D. W. Gordon, and J. A. Tuttle. Storage System Metrics for Evaluating Disk Array Organization.
 M. Y. Kim. Synchronized Disk Iterleaving. IEEE Trans. on Computers, pages 978–988, November 1986.
 M. S. Lakshmi and P. S. Yu. Effect of skew on join performance in parallel architectures. In Int. Symposium on
Databases in Parallel and Distributed Systems, pages 107–120, 1988.
 E. Lee. Software and Performance Issues in the Implementation of a RAID Prototype. May 1990.
 E. Lee and R. Katz. Performance Consequences of Parity Placement in Disk Arrays. pages 190–199, 1991.
 J. Menon and J. Cortney. TheArchitecture of a Fault-Tolerant Cached RAID Controller. In 20th Annual International
Symposium on Computer Architecture, pages 76–86, San Diego, CA, May 1993.
 J. Menon and J. Kasson. Methods for Improved Update Performance of Disk Arrays. Proceedings of the Hawaii
International Conference on System Sciences, pages 74–83, 1992.
 J. Menon and D. Mattson. Comparison of Sparing Alternatives for Disk Arrays. Proceedings of the International
Symposium on Computer Architecture, 1992.
 A. Merchant and P. S. Yu. Design andModeling of Clustered RAID. Proceedings of the International Symposium on
Fault-Tolerant Computing, pages 140–149, 1992.
 Richard R. Muntz and John C.S. Lui. Performance analysis of disk arrays under failure. VLDB Conference, pages
 David A. Patterson, Garth Gibson, and Randy H. Katz. A Case for Redundant Arrays of Inexpensive Disks (RAID).
ACM SIGMOD Conference, pages 109–116, 1988.
 M. Schulze, G. Gibson, R. Katz, and D. Patterson. How-Reliable is a RAID? COMPCON, pages 118–123, 1989.
 D. Stodolsky, G. A. Gibson, and M. Holland. Parity Logging, Overcoming the Small Writes Problem in Redundant
Disk Arrays. In 20th Annual International Symposium on Computer Architecture, pages 64–75, San Diego, CA,May
 M. Stonebraker. A Case for Shared Nothing. Database Engineering, 9(1), 1986.
 M. Stonebraker and G. A. Schloss. Distributed RAID - A New Multiple Copy Algorithm. Sixth Int’l. Conf on Data
Engineering, pages 430–437, 1990.
 D. Towsley, S. Chen, and S. P. Yu. Performance analysis of a fault tolerant mirrored disk system. Proceeding of
Performance ’90, pages 239–253, 1990.
 Philip S. Yu and Asit Dan. Effect of system dynamics on coupling architectures for transaction processing. Technical
Report RC 16606, IBM T.J. Watson Research Division, Feb 1991.
 Philip S. Yu and Asit Dan. Impact of affinity on the performance of coupling architectures for transaction processing.
Technical Report RC 16431, IBM T.J. Watson Research Division, Jan 1991.
“Disk Cooling” in Parallel Disk Systems
Dept. of Electr.. and Computer Sc.
Evanston, IL 60208
Dept. of Computer Science
University of Saarbrucken
D-66041 Saarbrucken, Germany
IBM Research Division
Almaden Research Center
650 Harry Road
San Jose, CA 95120
Paralleldisk systemsprovide opportunitiesfor high performance I/O by supportingefficiently intra-request
and inter-request parallelism. We review briefly the components of an intelligent file manager that performs
striping on an individual file basis and achieves load balancing by judicious file allocation and
dynamic redistribution of the data. The main part of the paper discusses our “disk cooling” procedure
for dynamic redistribution of data which is based on reallocation of file fragments. We show that this
heuristic method achieves excellent load balance in the presence of evolving access patterns. We report
on two sets of experiments: a synthetic experiment which exhibits a self-similar skew in the data access
patterns and a trace-based experiment where we study the impact of the file fragment size on the cooling
Parallel disk systems are of great importance tomassively parallel computers since they are scalable and they can
ensure that I/O is not the limiting factor in achieving high speedup. However, to make effective use of commercially
available architectures, it is necessary to develop intelligent software tools that allow automatic tuning of
the parallel disk system to varying workloads. The choice of a striping unit and whether to choose a file-specific
striping unit are important parameters that affect the response time and throughput of the system. Equally important
are the decisions of how to allocate the files on the disks and how to perform redistribution of the files when
access patterns change.
We have developed an intelligentfilemanager, called FIVE, for parallel disk systems that can performstriping
on a file-specific or global basis, as desired by the application, and in addition it achieves load balancing by judicious
file allocation and dynamic redistribution of data. Our systemis geared toward software-controlled parallel
disk systems in which each disk can be accessed individually. The system has the following salient properties:
It consists of modular blocks that can be invoked independently; in particular the algorithms for file allocation
and redistribution of data can be used regardless of whether striping is employed or not.
The research of this author was partially supported by NASA-Ames grant NAG2-846 and by NSF grant IRI-9303583.
This work was performed while the author was at ETH Zurich, Switzerland.
It uses simple but effective heuristics that incur only little overhead.
Its constituent algorithms can be invoked on-line, i.e., concurrently with regular requests to existing files.
The heuristics for data placement and redistribution of data can be integrated with the fault tolerance techniques
developed for RAIDs as well as various forms of data mirroring.
In this paper we discussmainly our “disk cooling” procedure for dynamic redistribution of data which is based on
reallocation of file fragments. We show that this heuristicmethod achieves excellent load balance in the presence
of evolving access patterns. In addition, we also discuss opportunities for fine-tuning our disk cooling procedure
so that the unit of reallocation can be chosen in order to account for the cost/benefit of the redistribution.
The remainder of this paper is organized as follows. In Section 2 we review briefly our file partitioning
method and discuss the relationship between partitioning and load balancing. In Section 3 we present our load
balancing procedure, concentrating on disk cooling and the bookkeeping steps required to keep track of its dynamically
changing statistics. Section 4 reports on two sets of experiments: a synthetic experiment which exhibits
a recursive skew of access and a trace-based experiment where we study the impact of the file fragment size on
the cooling procedure. Section 5 concludes with a brief discussion of issues under investigation.
2 File Partitioning
File striping or declustering [10, 15] is a technique for file organization that divides a file into runs of logically
consecutive data units, called “stripes” which are then spread across a number of disks in order to reduce the
transfer time of a single request or to improve the throughput of multiple requests. The striping unit denotes
the number of logically consecutive data bytes or blocks stored per disk, and the degree of declustering of a file
denotes the number of disks over which a file is spread. In virtually all disk architectures that have been proposed
so far, the striping unit is chosen globally . This approach is suitable for scientific applications or pure on-line
transaction processing, in which all files have approximately the same sort of access characteristics (i.e., only
large requests or only single-block requests). However, as we have shown in  for many applications which
exhibit highly diverse file access characteristics (e.g., VLSI design, desktop publishing, etc.) it is desirable to
tune the striping unit individually for each file.
We have developed an analytic model for an open queueing system in order to determine heuristically the
optimal striping unit on an individual file basis or on a global basis [17, 20, 22]. We observe here that an open
queueing model is much more realistic for an environment where a large number of users issue requests to a
parallel disk system, as compared to the closed queueing model used in [3, 12] where the number of concurrent
I/O requests in the systemis fixed. In our system, the striping unit is chosen in terms of data blocks. Our striping
procedure can be applied to a file system in two different ways:
1. The striping unit of each file is chosen individually, based upon the file’s average request size R. Two
further options exist here. For low arrival rates of requests, where we can assume that no queueing delays
occur, the response time can be computed as if the system operates in single user mode. For higher loads,
the response time can be optimized by taking into account explicitly the arrival rate, , in addition to the
2. The striping unit can be determined globally by any of the two optionsmentioned above based on the overall
average request size R?.
Although the problems of file striping and load balancing are orthogonal issues, they are not completely independent.
In order to derive analytically the optimal striping unit, it is assumed in [17, 20, 22] that the system
is load balanced.
If the striping unit is a byte and all files are partitioned across all disks in the system, then we obtain a perfectly
balanced I/O load. In general, very small striping units lead to a good load balancing. But throughput
considerations require for many applications that we choose large striping units (e.g., the size of a cylinder). For
example, the parity striping scheme proposed in  is based on very large (possibly infinite) striping units, and
 proposes choosing both a small and a large striping unit for replicated data to support both on-line transaction
processing and decision-support queries as well as batch processing. However, a coarser striping unit increases
the probability of load imbalance under a skewed workload . In general, we can see that file striping can
help towards achieving good load balancing, but partitioning by itself is not sufficient for this goal. Additional
methods for load balancing are called for, regardless of whether data partitioning is used or not.
3 Load Balancing
The load balancing component of our intelligent file system consists of two independent modules: one that performs
file allocation  and the second that performs dynamic redistribution of data .
After the decision has been made to decluster a file over a particular number of disks, all striping units of
the file that are to be mapped to the same disk are combined into a single allocation unit called an extent. The
file allocation problem in a parallel disk system involves making a judicious decision about the disks on which
to place the extents so as to optimize the load. While this problem is similar to the file allocation problem in
distributed systems, it presents an additional constraint due to the need to consider also intra-request parallelism.
This implies, that not all extents of a file should be allocated to the same disk if intra-request parallelism is to be
In order to performthese load balancing procedures, i.e., file allocation and file redistribution, our file system
keeps track of the following related statistics :
the heat of extents (or alternatively, of the smallest units of data migration) and disks, where the heat is
defined as the number of block accesses of an extent or disk per time unit, as determined by statistical
observation over a certain period of time,
and the temperature of extents, which is defined as the ratio between heat and size.
3.1 File Allocation
A number of heuristic methods have been proposed for file allocation, with the simplest one being the roundrobin
scheme. A simple but effective heuristic algorithm for static file allocation, where all files are allocated
at the same time and the heat of each extent can be estimated in advance, has been introduced in  (see, for
example,  for a more sophisticated approach to statically load balanced data placement). The algorithmfirst
sorts all extents by descending heat and the extents are allocated in sort order. For each extent allocation, the
algorithm selects the disk with the lowest accumulated heat among the disks which have not yet been assigned
another extent of the same file.
We have extended this greedy algorithm in order to deal with dynamic file allocation . Since our algorithm
makes no assumptions about the heat of a new file at allocation time, the sorting step is eliminated and the
algorithm only uses the information about the heat of the files which have been allocated already and for which
statistics are available. The disk selection can be made in such a way as to consider also, if so desired, the cost
of additional I/Os necessary to perform partial disk reorganization. Partial disk reorganization may have to be
performed if, due to file additions and deletions, there is room to store an extent on a disk but the space is not
contiguous. Even more expensive is the situationwhen disk i has the lowest heat and may appear as the obvious
choice to store a new extent of a file, but this disk does not have enough free space. In order to make room for the
new extent we have to migrate one or more extents to a different disk. In order to account for these reorganization
costswe associatewith every disk a status variablewith regard to the extent chosen for allocation. The status
Input: D - number of disks
Hj - heat of extent j
i - heat of disk i
H - average disk heat
Ei - list of extents on disk i sorted in descending temperature order
D - list of disks sorted in ascending heat order
Step 0: Initialization: target = not found
Step 1: Select the hottest disk s
Step 2: Check trigger condition:
if Hs > H (1 + ) then
Step 3: while (Es not exhausted) and (target == not found) do
Select next extent e in Es
Step 4: while (D not exhausted) and (target == not found) do
Select next disk t in D in ascending heat order
if (t does not hold an extent of the file to which e belongs)
and STATUS(t) == FREE then
target = found
Step 5 : if s has no queue then
s = H?
t = H?
t < He then
reallocate extent e from disk s to disk t
update heat of disks s and t:
s = H?0
t = H?0
Figure 1: Basic disk cooling algorithm
variable can take the values FREE, FRAG and FULL, depending upon whether the disk (1) has enough free space
for the extent, (2) has enough space but the the space is fragmented or, (3) does not have enough free space. Our
file allocation algorithm has the option of selecting disks in increasing heat order without regard to their status.
Alternatively, we may select the disks in multiple passes, where in the first pass we only choose those that have
3.2 The “Disk Cooling” Procedure
In order to perform dynamic heat redistributionwe employ in our system a dynamic load balancing step, called
disk cooling. Basically, disk cooling is a greedy procedure which tries to determine the best candidate, i.e., extent,
to remove from the hottest disk in order to minimize the amount of data that is moved while obtaining the
maximal gain. The temperature metric is used as the criterion for selecting the extents to be reallocated, because
temperature reflects the benefit/cost ratio of the reallocation since benefit is proportional to heat (i.e., reduction of
heat) and cost is proportional to size (of the reallocated extents). Our basic disk cooling procedure is illustrated
in Figure 1. The extent to be moved, denoted by e, is reallocated on the coolest disk, denoted by t, such that t
does not hold already an extent of the corresponding file and t has enough contiguous free space.
In our system the disk cooling procedure is implemented as a background demon which is invoked at fixed
intervals in time. The procedure checks first if the trigger condition is satisfied or not (Steps 1 and 2 in Figure 1).
If the trigger condition is false, the systemis considered load balanced and no cooling action is performed. In the
basic disk cooling procedure the system is not considered load balanced if the heat of the hottest disk exceeds the
average disk heat by a certain quantity . It is important to observe that during each invocation of the procedure
different disks can be selected as candidates for cooling after each cooling step.
Our procedure considers implicitly the cost/benefit ratio of a considered cooling action and only schedules it
for execution if is considered beneficial. These cost considerations are reflected in Step 5 of the algorithm. The
hottest disk is likely to have already a heavy share of the load, which we can “measure” by observing if its queue
is non-empty. A cooling action would most likely increase the load imbalance if a queue is present at the source
disk since it implies additional I/Os for the reorganization process. Hence, we choose not to schedule the cooling
action if this condition is satisfied. We also consider the cooling move not to be cost-beneficial if, would it be
executed, the heat of the target disk would exceed the heat of the source disk. Hence, although our background
demon is invoked a fixed number of times, only a fraction of these invocations result in data migration.
Our generic disk cooling procedure can be generalized in a number of ways. In  we have shown how
an explicit objective function based on disk heat variance (DHV) can be used in a more general test for the
cost/benefit of a cooling action. Thus, the benefit is computed by comparing the DHV after the potential cooling
step with the DHV before the potential cooling step. In addition, we can consider also explicitly the cost of performing
the cooling. Thus, a more accurate calculation of benefit and cost would consider not only the reduction
in heat on the origin disk and the increase in heat on the target disk, but also the additional heat caused by the reorganization
process itself. The cooling process is executed during two intervals of time, the first corresponding
to the read phase of the action and the second corresponding to the write phase of the action. The additional heat
generated during these phases can be computed by dividing the size of the extent to bemoved by the corresponding
duration of the phase. The duration times of the read and write phase of a cooling action can be estimated by
using a queueing model, as shown in .
Our disk cooling procedure can be fine-tuned so that the unit of reallocation is chosen dynamically in order
to increase the potential of a positive cost/benefit ratio. In the basic procedure given in Figure 1 the unit of
redistribution is assumed to be an extent. However, in the case of large extents that are very hot the cost of a
redistribution may be prohibitive. In this case, we can subdivide further an extent into a number of fragments
and use a fragment as the unit of redistribution. Since all fragments of an extent are of the same size we can now
base the choice of the migration candidates (see Step 3 in Figure 1) on the heat statistic instead of temperature.
In addition, the increase in the number of allocation units of a file also requires that we remove the allocation
constraint on the target disk, namely we do not require anymore that the disk should hold only one fragment of
a file. Hence, we put here the objective of a balanced load above the requirement that the degree of declustering
3.3 Heat Tracking
The dynamic tracking of the heat of blocks is implemented based on a moving average of the interarrival time
of requests to the same block. Conceptually, we keep track of the times when the last k requests to each block
occurred, where k is a fine-tuning parameter (in the range from5 to 50). To illustrate this bookkeeping procedure,
assume that a block is accessed at the points of time t1, t2,: : : , tn (n > k). Then the average interarrival time of
the k last requests is tntnk+1
k , and the estimated heat of the block is the corresponding reciprocal k
Upon the next access to this block, say at time tn+1, the block heat is re-estimated as k
One may conceive an alternative method for heat tracking that keeps a count of the number of requests to a
block within the last T seconds, where T would be a tuning parameter. The problem with such a request-count
approach is that it cannot track the heat of both hot and cold blocks in an equally responsivemanner. Hot blocks
would need a relatively short value of T to ensure that we become aware of heat variations quickly enough. Cold
blocks, on the other hand, would need a large value of T to ensure that we see a sufficient number of requests to
smooth out stochastic fluctuations. Themoving-averagemethod for the interarrival time does not have this problem
since a fixed value of k actually implies a short observation time window for hot blocks and a long window
for cold blocks. Moreover, extensive experimentation with traces from real applications with evolving access
patterns has shown that our tracking method works well for a wide spectrum of k values; the heat estimation is
fairly insensitive to the exact choice of k .
The adopted heat tracking method is very responsive to sudden increases of a block’s heat; the new access
frequency is fully reflected in the heat estimate after k requests, whichwould take only a shortwhile for hot blocks
(and reasonable values of k). However, themethod adapts the heat estimatemore slowlywhen a block exhibits a
sudden drop of its heat. In the extreme case, a hot block may suddenly cease to be accessed at all. In this case, we
would continue to keep the block’s old heat estimate as there are nomore newrequests to the block. To counteract
this form of erroneous heat estimation, we employ an additional “aging” method for the heat estimates. The
aging is implemented by periodically invoking a demon process that simulates “pseudo requests” to all blocks.
Whenever such a pseudo request would lead to a heat reduction, the block’s heat estimate is updated; otherwise
the pseudo request is ignored. For example, assume that there is a pseudo request at time t0 and consider a block
with heatH. We compute tentatively the new heat of the block as H0=k
, but we update the heat bookkeeping
only if H0 < H. The complete heat tracking method is illustrated in Figure 2.
to all blocks
Figure 2: Illustration of the heat tracking method for k = 3. The relevant interarrival times are shown by the
The described heat tracking method requires a space overhead per block of k+1 times the size of a floatingpoint
number. Since we want to keep this bookkeeping information in memory for fast cooling decisions, it is
typically unacceptable to track the heat of each individual block. For low-overhead heat tracking, we actually
apply the heat estimation procedure to entire extents (or fragments of a specified size). We keep track of the
times tn,: : : , tnk+1 of the last k requests that involve any blocks of the extent in the manner described above,
and also we keep the number of accessed blocks within the extent for each of the last k requests. Assume that
the average number of accessed blocks is R. Then the heat of the extent is estimated by kR
. Finally, we
estimate the heat of a fraction of an extent by assuming that each block in the extent has the same heat (which
is extent heat divided by extent size). This extent-based heat tracking method provides a compromise between
the accuracy and the space overhead of the block-based estimation procedure. The method has proven to be
sufficiently accurate in all our experimental studies (including studies with application traces).
4 Performance Studies
In this section we present an experimental evaluation of our dynamic disk cooling procedure. In order to study
the robustness of our procedure we performed two different sets of experiments. The first set of experiments
was based on a synthetic workload, enabling us to study systematically the effect of changes in arrival rates and
various patterns in data access skew. For the second set of experiments we used I/O traces from file servers at
the University of California.
4.1 The Performance Testbed
Our testbed consists of a load generator, the file system prototype FIVE, and a simulated I/O systemthat is using
CSIM . FIVE allows for striping of files on an individual basis and incorporates heuristics for file striping,
allocation, and dynamic load balancing. For the experiments presented here we wanted to study only the impact
of the various parameters relevant to the disk cooling procedure. Hence, we performed striping on a global basis,
using one track as the striping unit and static file allocation was performed using round-robin. The disk cooling
procedure is invoked as a background process running concurrently with regular requests to existing files. FIVE
can manage real data on real hardware or synthetic/real data on a simulated I/Osystem. For these experimentswe
used a simulated disk farm consisting of 32 disks whose parameters are described in Table 1. Note, that although
average figures are given for the seek time and rotational latency, our detailed simulation computes the actual
figures for each request by keeping track of the cylinder position of each disk arm.
capacity per disk [MBytes] 540 avg. seek time [ms] 12
# cylinder per disk 1435 avg. rotational latency [ms] 6.82
# tracks per cylinder 11 transfer rate [MBytes/s] 2.4
capacity per track [blocks] 35 block size [Bytes] 1024
# disks 32 total capacity [GBytes] 17
Table 1: Characteristics of the simulated disk farm
We performed a number of synthetic experiments in order to study systematically the effects of different parameters
on the disk cooling procedure. While experiments with real-life traces are important, it is often difficult to
obtain long enough traces that exhibit all the relevant patterns of access. The purpose of the synthetic experiments
was to study the impact of various arrival rates, degree of skew in data access, as well as fluctuations over
time in the skew.
For these experiments we used N = 1000 files each having 70 blocks (2 tracks). Each file resided on two
disks (i.e., the global striping unit was 1 track). Furthermore, as the file size was rather small, we considered
only entire files as migration candidates and did not investigate smaller fragment sizes. Each (read or write)
request accessed an entire file.1 Note that this synthetic setup captures the most important workload features of
applications such as office document filing, desktop publishing, etc., as far as load balancing issues are concerned.
In a real application, there would probably be more files and also larger files, but, on the other hand, I/O requests
would often access only a fraction of a file and a large fraction of the files would be accessed only extremely
infrequently. So we disregard the non-essential details for the sake of simplifying the experiments.
The I/O requests were generated so as to exhibit a self-similar skew in the data access pattern across the files
. We use an X/Y distribution of access, where X % of the requests are addressed to Y % of the files. Thus, if
the files are numbered from 1 to N, for a given setting of the parameters X and Y, the probability of accessing a
file numbered i, with i s, is given by the formula of :
Prob(i s) = s
In order to experiment with fluctuations in patterns of access, we have implemented a uniform shift procedure
which allows us to switch among the files the heats assigned to them from one period to the other. Let us assume
that during simulation period i the files numbered 1 through N have been assigned heats in decreasing order of
1In Section 4.3 we discuss experiments with larger file sizes and variable request sizes.
0 10 20 30 40 50
(a) avg. response time
0 10 20 30 40 50
(b) Cooling process statistics
0 10 20 30 40 50
(c) avg. response time
0 10 20 30 40 50
(d) Cooling process statistics
queue at source
no migration candidate
# of cooling steps
Figure 3: Average response time for synthetic experiments. skew 70/30. = 200. (a) and (b) one shift, (c) and
(d) three shifts.
magnitude; thus file numbered 1 was assigned the highest heat, h1, and so on, with the file numbered N being
assigned the smallest heat, hN. A shift in heat of magnitude r means that in the next simulated period, namely
i+ 1, the highest heat h1 is assigned to file numbered 1+ r, h2 is assigned to file numbered 2+ r, and so on in
In all experiments the simulation time was divided into 50 intervals. We report here on two sets of experiments
using a different degree of skew: the first one uses a 80/20 skew in access, while the second one uses a
70/30 skew. For each set of experiments we experimented with different shifts in skew: no shift for the entire
simulation versus one or three shifts in skew. In the case of one shift the magnitude chosen was r = 500 and the
shift occurred in the middle interval (number 25). In the experiments with three shifts the magnitude of the shift
was 250 and the shifts occurred during intervals 12, 25 and 37, respectively. Different arrival rates were used for
each set of experiments, but due to lack of space we limit ourselves here to only one arrival rate per X/Y skew.
Note that a lower skew in data access can handle a much higher arrival rate. This is due to the fact that for high
degrees of skew the vanilla round-robin allocation method thrashes above a certain arrival rate.
Figure 3 depicts the average response times and the cooling frequencies as they vary over the simulated time
period for a skew of 70/30. Figure 4 repeats these measurements for a skew of 80/20. The disk cooling procedure
achieves tremendous savings in the response time due to better load balancing, and hence reduced queueing
The histograms illustrated in the Figures depict the frequency of the datamigration steps invoked by our cooling
procedure, varying over the simulation intervals. The disk cooling procedure is implemented as a background
demon which is invoked a fixed number of times (i.e., 100 times) during each simulation period. The histograms
illustrate how many of these invocations actually resulted in data migrations being executed (cooling steps). An
invocation will not result in a cooling action if the system determines that the cost/benefit ratio is not favorable.
The cases when no cooling actions occur are divided into two categories in our histograms: queue at source and
no migration candidates. Thequeue at source category counts those invocations where a queue is observed at the
source disk. The no migration candidate category includes those invocations where (1) all extents on the hottest
disk are so hot that after a move the target disk would become hotter than the source disk before the move, or (2)
all remaining extents have no observed I/Os. The first case was discussed already in Section 3. The second case
is related to the fact that our disk cooling procedure has no a-priori knowledge about the heat of any of the extents.
Hence, initially the heat of any extent is assumed to be zero and the disk cooling procedure is not executed
further when we reach extents with no observed I/Os.
0 10 20 30 40 50
(a) avg. response time
0 10 20 30 40 50
(b) Cooling process statistics
0 10 20 30 40 50
(c) avg. response time
0 10 20 30 40 50
(d) Cooling process statistics
queue at source
no migration candidate
# of cooling steps
Figure 4: Average response time for synthetic experiments. Skew 80/20. = 95
. (a) and (b) one shift, (c) and
(d) three shifts.
In all experiments our algorithm initiated a larger number of cooling steps during the initial learning phase.
Figure 4 shows that in the case of a 80/20 skew, the number of cooling steps initiated subsequently was rather
small. In addition, the experiments illustrate that our disk cooling procedure reacts fast to changes in the access
pattern. As observed in Figure 4, a sudden shift during interval 25 caused a slight increase in the average response
time. But the system is fast to recognize it, and immediately in the next time interval cooling actions are taken
to readjust the load balance. The experiments with three shifts show a similar pattern: very few and singular
cooling actions, performed after the shifts occurs. On the other hand, the experiments depicted in Figure 3 show
a different configuration. In these experiments we observe that the cooling steps occur continuously throughout
the entire simulation period. This is due to the fact that for a more moderate skew the disk cooling procedure
has more degrees of freedom for the migration. With a high skew it often happens that we end up with one or
two disks which have only one extent, namely the hottest ones. When this occurs, no suitable target disk can be
found for the migration since its heat will now become the bottleneck.
4.3 Trace-based Workload
For this experimental study we used a 48-hours trace fromthe University of California, Berkeley, a period during
which the majority of the applications dealt with VLSI design and simulation. The original trace, described in
, was reformatted in order to feed it into our testbed. Furthermore, we removed all short-lived files with a lifetime
of less than a minute, assuming that these files would be cached on a permanent basis. An important feature
of this trace is the constantly changing pattern of access to the individual files. Files accessed frequently at the
beginning of the trace are hardly requested anymore at the end of the trace. Hence, we are dealing here with a
skewed access distribution that is undergoing continual shifts. In addition, the file sizes are substantially larger
than in the synthetic experiment. This enabled us to study here the impact of the file fragment size on the disk
The trace consists of approximately 440,000 requests for data in 14,800 files. The average request size is
about 107 KBytes, but a wide variance in the request sizes is exhibited. The original average arrival rate of the
trace was = 2:45 requests per second; we accelerated the load by a “speed-up factor” of 10 for a stress test,
thus the effective average arrival rate was = 25 requests per second. Note, however, that the trace contained
heavy load surges with much higher arrival rates. All files were striped with a striping unit of one cylinder (385
blocks) and allocated on the disks in a round-robin manner.
0 10 20 30 40 50
(a) avg. response time
0 10 20 30 40 50
(b) Cooling process statistics
queue at source
no migration candidate
# of cooling steps
Figure 5: Response time for the trace-driven experiment
As in the synthetic experiments, the disk cooling procedure had no a-priori information about the access frequencies
of the individual extents or fragments. The heat statisticswere collected and updated dynamically using
a moving window containing the last k = 20 requests to a given extent (fragment).
In our first experiment we assumed that the fragment size for the reallocation procedure is 210 Kbytes (six
tracks). Figure 5 (a) shows the average response time varying over time obtained by using our disk cooling procedure
versus a vanillamethod that does notmake use of disk cooling. As expected, our load balancing algorithm
exhibits a learning phase at the beginning of the simulation, during which it collects heat statistics, which corresponds
to the peak in access time of Figure 5 (a). After this the average response time obtainedwith disk cooling
drops substantially. Over the entire simulation period the average response time measured was 1.085 seconds
without disk cooling, versus 0.297 seconds with disk cooling.
Figure 5 (b) illustrates the frequency of the datamigration steps invoked by our load balancing procedure. We
observe that the cooling steps are executed throughoutthe entire simulation period due to the continually evolving
pattern of access, i.e., the constant shift in skew. Overall, 1798 cooling steps were executed each requiring on the
average 0.13 seconds. Between simulation periods 17 and 28, and then between periods 40 and 49, the cooling
quiescents somewhat. This is due to the fact that the trace covers a 48 hours period and these periods correspond
to the two lightly loaded night intervals (as backup activity was not recorded in the trace).
In order to study the impact of file fragment size on the disk cooling procedure we designed a further set of
experiments. The file fragment size was varied across experiments from FG=0, (no fragmentation—the extent
is moved in its entirety) to FG =35 KBytes (one track). For each experiment, i.e., simulation period, the fragment
size was kept fixed. Figure 6 shows the average response times with disk cooling for different fragment
sizes. We observe that in the extreme case where no fragmentation is used (FG=0) the disk cooling procedure
performs almost identically to the vanilla round-robin allocation method (see Figure 5 (a) for comparison). This
phenomenon is due to two factors characteristic to large files. First, the extents can become so large that the cost
of the migration exceeds the benefit of the move. Secondly, the number of extents in a file can be quite large,
hence it becomes difficult to find a target disk which satisfies the constraint that no other extents of the file are
already stored there. At the other extreme setting of the fragment size, i.e., FG = 35 Kbytes (one track) we also
observed no improvement versus the vanilla algorithm which does not perform disk cooling, since the benefits
of each migration are too small. However, for fragment sizes of 140 Kbytes, 210 Kbytes or 385 Kbytes (one
cylinder) we observed a substantial improvement of the response time. All these fragment sizes offer a good
compromise between costs and benefit of the redistribution. Furthermore, it is worthwhile to observe that the
0 10 20 30 40 50
Figure 6: Response time for different fragment sizes
response times are fairly insensitive to the exact settings of the fragment size.
The trace-based experiments confirm the results obtained in the synthetic experiments. The cooling procedure
is robust and fine-tunedwith respect to changes in access patterns. If the pattern of access changes at certain time
intervals and remains relatively stable in between, as was the case in the 1 shift/ 3 shifts experiments, the load
balance can be restored with very few cooling steps if the skew is large. On the other hand, for smaller values
of data skew, or for a constant shift in skew, as happened with the trace, the migration activities occur with high
frequency. In all cases, the migration is executed only if the benefits of cooling exceed the cost of the move and
this trade-off presents opportunities for fine-tuning the choice of the migration unit (i.e., the fragment size).
5 Further Issues
The disk coolingmethod that we have discussed in detail in this paper forms a major component in an integrated
butmodular set of data placement and dynamic reorganization heuristics that we have developed for parallel disk
systems. This package is implemented in the FIVE experimental file system. We are in the process of investigating
additional dimensions of the cooling method and exploring various generalizations of our approach.
Most importantly, we are considering the impact of caching on our heat tracking method. For this purpose,
we distinguish the logical heat and the physical heat of a block, where the former includes all accesses to a block
and the latter counts only those accesses that incurred disk I/O.While it may seem at a first glance that disk load
balancing needs to consider only physical heat, such an approach would be very sensitive to sudden changes in
the cache effectivity. For example, sudden load imbalances may result from a decision of the cache manager to
drop several hot blocks that happen to reside on the same disk. This could occur when the cache manager needs
to set aside a large amount of working memory (e.g., for a hash join), which effectively changes the memory
size as far as disk block caching is concerned. On the other hand, simply balancing the logical heat of the data
does not necessarily achieve the goal of disk load balancing, since different disks may have different quantities
of “cached heat” (i.e., accesses that are serviced by the cache).
Another important issue that needs further investigation is the synchronization and fault tolerance of the various
reorganization steps that wemay invoke on-line in our system(e.g., the cooling steps). Basically, this requires
critical sections for accesses to the block addressing tables and also logging the changes to these tables. We are
currently working out details of these issues. Note that none of the on-line reorganizations requires holding locks
on the actual data blocks for transaction-length or longer duration.
Finally, we are aiming to generalize our approaches to data partitioning, data allocation, and dynamic load
balancing to arbitrary distributedstorage systems such as shared-nothingparallel computers orworkstation farms.
While basic principles of our approach can be carried over, differences in the various cost metrics demand us to
reconsider the benefit/cost optimization. We have also started studying a broader spectrum of application workloads.
In particular, we want to address the performance requirements of multimedia applications with accesses
to continuous, delay-sensitive media like audio and video. Although there exists some promising work on this
subject (e.g., [2, 6, 9]), we believe that substantiallymore research is needed towards efficient and cost-effective
multimedia data management.
 Baker, M.G., Hartman, J.H., Kupfer, M.D., Shirriff, K.W., and Ousterhout, J.K., “Measurements of a Distributed File
System,” Proc. 13th ACM Symposium on Operating Systems Principles, 1991.
 Chiueh, T., and Katz, R., “Multi-ResolutionVideo Representation for Parallel Disk Arrays,” Proc. ACMMultimedia
 Chen, P.M. and Patterson, D.A.,“ Maximizing Performance in a Striped Disk-Array,” Proc. 17th Int. Symposium on
Computer Architecture, 1990.
 Chen, P.M., Lee, E.K., Gibson, G.A., Katz, R.H., and Patterson, D.A., “RAID: High-Performance, Reliable Secondary
Storage,” Technical Report UCB/CSD-93-778, Department of Computer Science, University of California
at Berkeley, 1993.
 Copeland, G., Alexander, W., Boughter, E., and Keller, T., “Data Placement in Bubba,” Proc. ACM SIGMOD Conf.,
 Gemmel, J., and Christodoulakis, S., “Principles of Delay-Sensitive Multimedia Data Storage and Retrieval,” ACM
Transactions on Information Systems, Vol. 10, No. 1, 1992.
 Gray, J.N., Horst B., and Walker, M., “Parity Striping of Disc Arrays: Low-Cost Reliable Storage with Acceptable
Throughput,” Proc. 16th Int. Conf. on Very Large Data Bases, 1990.
 Gray, J., Sundaresan, P., Englert, S., Baclawski, K., and Weinberger, P.J., “Quickly Generating Billion-Record Synthetic
Databases,” Proc. ACM SIGMOD Int. Conf., 1994.
 Ghandeharizadeh, S., and Ramos, L., “Continuous Retrieval ofMultimedia Data Using Parallelism,” IEEE Transactions
on Knowledge and Data Engineering, Vol. 5, No. 4, 1993.
 Kim, M.Y.,“Synchronized Disk Interleaving,” IEEE Transactions on Computers, Vol. C-35, No. 11, 1986.
 Knuth, D.E.,“The Art of Computer Programming. Vol. 3: Sorting and Searching,” Addison-Wesley, 1973.
 Lee, E.K., andKatz, R.H., “AnAnalytic Performance Model ofDiskArrays,” Proc. ACMSIGMETRICSConf., 1993.
 Livny, M., Khoshafian, S., and Boral, H., “Multi-DiskManagement Algorithms,” Proc. ACM SIGMETRICS Conf.,
 Merchant, A. and Yu, P.S.,“Performance Analysis of a Dual Striping Strategy for Replicated Disk Arrays,” Proc. 2nd
Int. Conf. on Parallel and Distributed Information Systems, 1993.
 Salem, K., and Garcia-Molina, H.,“Disk Striping,” Proc. 2nd Int. Conf. on Data Engineering, 1986.
 Scheuermann, P., Weikum, G., and Zabback, P., “Adaptive Load Balancing in Disk Arrays,” Proc. 4th Int. Conf. on
Foundations of Data Organization and Algorithms, Lecture Notes in Computer Science, No. 730, 1993.
 Scheuermann, P., Weikum, G., and Zabback, P., “Data Partitioning and Load Balancing in Parallel Disk Systems,”
Technical Report 209, Department of Computer Science, ETH Zurich, January 1994.
 Schwetman, H., “CSIM Reference Manual (Revision 16),” Technical Report ACA-ST-252-87, MCC, 1992.
 Weikum, G., Zabback, P., and Scheuermann, P., “Dynamic File Allocation in Disk Arrays,” Proc. ACM SIGMOD
Int. Conf., 1991.
 Weikum, G., and Zabback, P., “Tuning of StripingUnits in Disk-Array-Based File Systems,” Proc. 2nd Int.Workshop
on Research Issues on Data Engineering: Transaction and Query Processing (RIDE-TQP), 1992.
 Wolf, J., “The Placement Optimization Program: A Practical Solution to the Disk File Assignment Problem,” Proc.
ACM SIGMETRICS Conf., 1989
 Zabback, P., “I/O Parallelism in Database Systems,” Ph.D. Thesis (in German), Department of Computer Science,
ETH Zurich, 1994.
Issues in Parallel Information Retrieval
Anthony Tomasic Hector Garcia-Molina
Stanford University, Stanford, CA 94305-2140
The proliferation of the world’s “information highways” has renewed interest in efficient document indexing
techniques. In this article, we provide an overview of the issues in parallel information retrieval.
To illustrate,we discuss an example of physical index design issues for inverted indexes, a common form
of document index. Advantages and disadvantages for query processing are discussed. Finally, to provide
an overview of design issues for distributed architectures, we discuss the parameters involved in the
design of a system and rank them in terms of their influence on query response time.
As the data volume and query processing loads increase, companies that provide information retrieval services
are turning to parallel storage and searching. The idea is to partition large document collections, and their index
structures across computers. This allows for larger storage capacities and permits searches to be in parallel.
In this article we sample research in the area of parallel information retrieval. We start by summarizing basic
information retrieval concepts, and then describe how they have been applied in a parallel environment. We also
give a short summary of our own research in this area, mainly as an example of the types of algorithms that need
to be developed, and the system issues that need to be studied.
2 Information Retrieval Basics
For an introduction to full-text document retrieval and information retrieval systems, see reference . An information
retrieval model (IRM) defines the interaction between a user and an information retrieval system and
consists of three parts: a document representation, a user need and a matching function.
The boolean IRM is provided by most existing commercial information retrieval systems. Its document representation
is the set of words that appear in each document. Typically, each word is also typed to indicate if it
appears in the title, abstract, or some other field of the document. The boolean IRM user need is represented by
a boolean query. A query consists of a collection of pairs of words and types structured with boolean operators.
For example the query title information and title retrieval or abstract inverted contains three pairs and two operators.
Thematching function of a query in the boolean IRMis boolean satisfiability of a document representation
with respect to the query.
The vector IRM is popular in academic prototypes for information retrieval systems and has recently gained
commercial acceptance. Its document representation is the set of words that appear in each document and an
associated weight with each word. The weight indicates the “relevance” of the corresponding word to the document.
Thus, a document is represented as a vector. A vector IRMuser need is represented by another vector (this
vector can be extracted from a document or a set of words provided by a user). The matching function computes
the similarity between the user need and the documents. Thus, all the documents can be ranked with respect to
the similarity. Typically, the topmost similar documents are returned to the user as an answer. There is much
research on the assignment of weights to words and on the effectiveness of various matching functions for information
retrieval. However, both the boolean IRM and the vector IRM and associated variation of these models
can be computed efficiently with inverted lists. (See Section 4 for a description of inverted lists.) Reference 
surveys information retrieval models.
The focus of traditional information retrieval research is to develop IRMs that provide the most effective interaction
with the user. Our focus in this article, however, is in providing the most efficient interaction with the
user in terms of response time, throughput and other measures, regardless of which IRM is used.
In the design of full-text document retrieval systems, there is a basic trade-off between the time to process
the document database and the time to process queries. Broadly speaking, the more time spent processing the
document database (i.e., building indexes) the less time is spent processing queries. In some scenarios (such as
government monitoring of communication), a tremendous amount of information must be queried by only a few
queries. In this case, time spent indexing is wasted and linear searching of documents is more efficient. Work in
this area concentrates on hardware processors for speeding up the scanning of text . More typically, indexing
the documents isworthwhilebecause the cost can be amortized acrossmany queries. We consider only these latter
Emrath’s thesis  explores this trade-off between query and update time by providing a data structure that
can be tuned in the amount of information indexed. Essentially, the database is partitioned into equal sized “pages.”
A page is a fixed number of words located together in a document. Duplicate occurrences of words are dropped
within a page. If the page is large, many duplicates are dropped from the index, speeding up indexing time. If
the page is small, few duplicate words are dropped, slowing down indexing time. For certain applications this
tuning of the data structure works well.
More recent work [18, 26, 27] uses physical index design to express the trade-off. The collection of documents
is partitioned and each partition has an independent index at the physical index design level, but the entire
collection has a single logical index. This provides fast update time but slow query time since each physical
index must be searched. To provide fast query time, the physical indexes are merged according to a family of
algorithms. More typically, indexing the documents in a single physical index is worthwhile because the cost can
be amortized across many queries. We consider only these latter systems for the remainder of this article.
Much research has gone into designing data structures for indexing text. Faloutsos  is a survey of this issue.
One approach is the use of signature schemes (also known as superimposed coding) . Here, each word is assigned
a random (hashed) k-bit code of an n-bit vector – for example the word “information” might correspond
to bit positions 10 and 20 of a 2 kilobyte vector. Each document is represented by an n-bit vector containing
the union of all the k-bit codes of all the words in the document. Queries are constructed by producing an n-bit
vector of the k-bit codes of the words in the query. Matching is performed by comparing a query against the document
vectors in the database. This scheme is used because the signatures of documents can be constructed in
linear time. Unfortunately, the matching process produces “false drops” where different words or combinations
of words are mapped into the same k-bit codes. One approach is to ignore false drops and inform the user that
some additional documents may be returned. We do not consider this approach further. Otherwise, each document
in the result of the matching process must be checked for false drops. While the number of false drops can
be statistically controlled for the average case, the worst-case behavior of this data structure implies checking every
document in the database for some queries, which is prohibitively expensive for large document collections.
Lin  describes a signature scheme where multiple independent signatures are used to control false drops and
to improve parallel performance.
Another data structure is PATRICIA trees and PAT arrays [9, 10]. Here, the database is represented as one
database string by placing documents end-to-end. A tree is constructed that indexes the semi-infinite strings of
the database string. A semi-infinite string is a substring of the database string starting at some point and proceeding
until it is a unique substring. The PAT system provides indexing and querying over semi-infinite strings.
The New Oxford English Dictionary has be index using this data structure. The query time, indexing time, and
storage efficiency are approximately the same as inverted lists. The techniques described here can be applied to
this data structure.
For commercial full-text retrieval systems, inverted files or inverted indexes [8, 13] are typically used. Note
that the information represented in each posting (each element of an inverted list) varies depending on the type
of information retrieval system. For a boolean IRM full-text information retrieval system, the posting contains
the document identifier and the position (as a byte offset or word offset from the beginning of the document)
of the corresponding word. For a boolean IRM abstracts text information retrieval system, the posting contains
the document identifier without a positional offset (since duplicate occurrences of a word in a document are not
represented in these systems). For a vector IRM full-text or abstracts information retrieval systems the posting
contains the document identifier and a weight. All of the above systems can be typed. In this case, the type
system can be encoded by setting aside extra bits in each posting to indicate which fields the word appears in the
document. Other methods of representing the type information are also used. As the information retrieval model
becomes more complicated, more information is typically placed in each posting.
A related area of research is the compression of inverted indexes [29, 30]. The inverted index for a full-text
information retrieval system is very large – typically on the same scale in size as the text. In fact, the original
documents (minus punctuation) can be reconstructed from the inverted index. Thus, one interesting physical
design issue is the impact of the compression ratio of the inverted index on response time. We return to this issue
in Section 6.
3 Parallel Query Processing
Various distributed and parallel hardware architectures can be applied to the problem of information retrieval. A
series of papers by Stanfill studies this problem for a ConnectionMachine. In reference , signature schemes
are used. A companion paper by Stone  argues that inverted lists on a single processor are more efficient.
In reference , inverted lists are used to support parallel query processing (in a fashion similar to that used
by the system index organization that will discussed in Section 4). Finally, in reference , an improvement
of the previous paper based on the physical organization of inverted lists is described. The technique essentially
improves the alignment of processors to data.
An implementation of vector IRM full-text information retrieval is described in reference  for the POOMA
machine. The POOMAmachine is a 100-node, 2-dmesh communication network where each node has 16MB of
memory and a processor. One out of five nodes has an ethernet connection and one half of the nodes have a local
disk. The implementation partitions the documents among the processors and builds a local inverted index of the
partition. (This approach is similar to the host index organization of Section 4; however there are two processors
per disk, as opposed to multiple disks per processor.) This paper cites a 2.098 second estimated query response
time for a 191-term query on a database of 136,020 documents with a 20-node machine.
Some preliminary experimental results are reported in reference  for a 16 processor farm (Meiko Computing
Surface). The vector IRM is used here and a signature scheme is used as the data structure. Unfortunately,
the database has only 6,004 documents and the query workload only 35 queries.
The performance of some aspects of query and update processing of an implementation of a boolean IRMfulltext
information retrieval is discussed in reference  for a symmetric shared-memory multiprocessor (Sequent).
Reference  presents a discussion of the architecture issues in implementing the IBMSTAIRS information
retrieval system on a network of personal computers. This paper argues for the physical distribution of inverted
lists across multiple machines when the size of a single database is larger than the storage capacity of a node on
the network. This idea is essentially a special case of disk striping, where an object (in this case an inverted list)
is partitioned across disks.
In the analysis of query processing, a query can be divided into three parts: parsing the query, matching the
query against the database, and retrieving the documents in the answer. Parsing consumes few resources and is
d 0 d 1 d 2 d 3
a b a b
BUS 0 BUS 1
CPU 0 CPU 1
Figure 1: A example set of four documents and an example hardware configuration.
typically the same for all information retrieval systems. Retrieving of documents offers some interesting issues
(such as placement of the documents) but again few resources are needed. Burkowski  examines the performance
problem of the interaction between query processing and document retrieval and studies the issue of the
physical organization of documents and indices. His paper models queries and documents analytically and simulates
a collection of servers on a local-area network.
Schatz  describes the implementation of a distributed information retrieval system. Here, performance
improvements come from changing the behavior of the interface to reduce network traffic between the client interface
and the backend information retrieval system. These ideas are complementary our work. Three improvements
are offered. First, summaries of documents (or the first page) are retrieved instead of entire documents.
This scheme reduces the amount of network traffic to answer an initial query and shortens the time to present
the first result of a query, but lengthens the time to present the entire answer. Second, “related” information such
as document structure definitions are cached to speed up user navigation through a set of documents. Third, the
contents of documents (as opposed to summaries) are prefetched while the user interface is idle.
Our own work [23, 24, 25] compares various options for partitioning an inverted list index across a sharednothingmulti-
processor. (Reference  considers shared-everything multi-processors.) Simulated query loads
are used in [24, 25], while  uses a trace-driven simulation.
4 Some Physical Design Choices
To illustratemore concretely the types of choices that are faced in partitioning index structures across machines,
in this sectionwe briefly describe the choices for an inverted-lists index, using the terminology of . As stated
earlier, this is the most popular type of index in commercial systems.
The left hand side of Figure 1 shows four sample documents, D0, D1, D2, D3, that could be stored in an
information retrieval system. Each document contains a set of words (the text), and each of these words (maybe
with a few exceptions) are used to index the document. In Figure 1, the words in our documents are shownwithin
the document box, e.g., document D0 contains words a and b.
As discussed in Section 1, full-text document retrieval systems traditionally build inverted lists on disk to find
documents quickly [8, 13]. For example, the inverted list for word b would be b: (D0,1), (D2,1), (D3,1). Each
pair in the list is a posting that indicates an occurrence of the word (document id, position). To find documents
containing word b, the system needs to retrieve only this list. To find documents containing both a and b, the
system could retrieve the lists for a and b and intersect them. The position information in the list is used to answer
queries involving distances, e.g., find documents where a and b occur within so many positions of each other.
Index Disk Inverted Lists in word: (Document, Offset) form
Host d 0 a: (D0, 0), (D1, 0)
d 1 b: (D0, 1)
d 2 a: (D2, 0), (D3, 0); c: (D3, 2)
d 3 b: (D2, 1), (D3, 1); d: (D3, 3)
System d 0 a: (D0, 0), (D1, 0), (D2, 0), (D3, 0)
d 1 b: (D0, 1), (D2, 1), (D3, 1)
d 2 c: (D3, 2)
d 3 d: (D3, 3)
Table 2: The various inverted index organizations for Figure 1.
Suppose thatwe wish to store the inverted lists on amultiprocessor like the one shown on the right in Figure 1.
This system has two processors (CPUs), each with a disk controller and I/O bus. (Each CPU has its own local
memory.) Each bus has two disks on it. The CPUs are connected by a local area network. Table 2 shows four
options for storing the lists. The host and I/O bus organizations are identical in this example because each CPU
has only one I/O bus.
In the system index organization, the full lists are spread evenly across all the disks in the system. For example,
the inverted list of word b discussed above happened to be placed on disk d1. This organization essentially
divides the keywords among the processors.
In the host index organization, documents are partitioned into two groups, one for each CPU. Here we assume
that documents D0, D1 are assigned to CPU 0, and D2, D3 to CPU 1. Within each partition we again build
inverted lists. The lists are then uniformly dispersed among the disks attached to the CPUs. For example, for
CPU 1, the list for a is on d2, the list for b is on d3, and so on.
Clearly, many choices are available for physical index organization beyond those described here. We cannot
consider all possible organizations. Our criteria for choosing these two organizations focuses first on the optimization
of queries as opposed to updates. Thus, we assume that the inverted lists on each machine are stored
contiguously on disk. Second, we are interested in the interaction between the physical index organization and
the allocation of resources (CPUs, disks, I/O buses) of a shared-nothing distributed system. In addition, we have
studied issues such as striping and caching of the physical index organization with respect to a single host.
5 Query Processing
Given a physical index partition like the ones illustrated in the previous section, how does one process queries?
To illustrate, let us focus on a particular type of query, a “boolean and” query. Such queries are of the form
a ^ b ^ c : : :, and find the documents containing all the listed words. The words appearing in a query are termed
keywords. Given a query a ^ b : : : the document retrieval system generates the answer set for the document
identifiers of all the documents that match the query. A match is a document that contains the words appearing
in the query.
Notice that boolean-and queries are themost primitive ones. For instance, amore complex search such as (a^
b) OR (c^d) can bemodeled as two simple and-queries whose answer sets aremerged. Adistance query “Find a
and b occurringwithin x positions” can bemodeled by the query a^b followed by comparing the positions of the
occurrences. Thus, the query processing strategies for the more complex queries can be based on the strategies
we will illustrate here for the simple boolean-and queries.
For the host index organization, boolean-and queries can processed as follows. The query a ^ b::: is initially
processed at a home site. That site issues subqueries to all hosts; each subquery contains the same keywords as
the original query. A subquery is processed by a host by reading all the lists involved, intersecting them, and
producing a list of matching documents. The answer set of a subquery, termed the partial answer set, is sent to
the home host, which concatenates all the partial answer sets to produce the answer set.
In the system index organization, the subquery sent to a given host contains only the keywords that are handled
by that host. If a host receives a query with a single keyword, it fetches the corresponding inverted list and
returns it to the home host. If the subquery contains multiple keywords, the host intersects the corresponding
lists, and sends the result as the partial answer set. The home host intersects (instead of concatenates) the partial
answer sets to obtain the final answer.
There are many interesting trade-offs among the storage organizations and query processing strategies. For
instance, with the system index organization, there are fewer I/Os. That is, the a list is stored in a single place
on disk. To read it, the CPU can initiate a single I/O, the disk head moves to the location, and the list is read.
(This may involve the transfer of multiple blocks). In the host index organization, on the other hand, the a list
is actually stored on, say, 4 processors. To read these list fragments, 4 I/Os must be initiated, four heads must
move, and four transfers occur. However, each of the transfers is roughly a fourth of the size, and they can take
place in parallel. So, even though we are consuming more resources (more CPU cycles to start more I/Os, and
more disk seeks), the list may be read more quickly.
The systemindex organizationmay save disk resources, but it consumes more resources at the network level.
Notice that in our example, the entire c list is transferred fromCPU 1 to CPU 0, and these inverted lists are usually
much longer than the document lists exchanged under the other schemes. However, the long inverted list transfers
do not occur in all cases. For example, the query “Find documents with a and b” (systemindex organization) does
not involve any such transfers since all lists involved are within one computer. Also, it is possible to reduce the
size of the transmitted inverted lists by moving the shortest list. For example, in our “Find documents with a and
c”, we can move the shorter list of a and c to the other computer.
It is also important to notice that the query algorithms we have discussed can be optimized in a variety of
ways. To illustrate, let us describe one possible optimization for the system index organization. We call this
optimization Prefetch I; it is a heuristic and in some cases itmay not actually improve performance. (Other query
optimization techniques have been studied in the literature.)
In the Prefetch I algorithm, the home host determines the query keyword k that has the shortest inverted list.
We assume that hosts have information on keyword frequencies; if not, Prefetch I is not applicable. In phase 1,
the home host sends a single subquery containing k to the host that handles k. When the home host receives the
partial answer set, it starts phase 2, which is the same as in the un-optimized algorithm, except that the partial
answer set is attached to all subqueries. Before a host returns its partial answer set, it intersects it with the partial
answer set of the phase 1 subquery, which reduces the size of the partial answer sets that are returned in phase 2.
6 Experimental Parameters
In this sectionwe summarize two studieswe have performed to evaluate the index partition and query processing
trade-offs. We believe they are representative of the types of analysis that needs to be performed to evaluate
physical design alternatives for information retrieval. In particular, we focus on the experimental parameters
used and their impact on response time. Our ranking of these parameters gives an overview on the important
areas to consider when designing an information retrieval system. In addition to the simulation work described
here, a general interest in the performance of text document retrieval systems has led to a standardization effort
for benchmarking of systems .
The first study  focused on full-text information retrieval. In full-text retrieval, the inverted index contains
essentially the same information as the documents, since the position of each word in each document is
recorded. Our inverted list model was based on experimental data, and our query model was based on a probabilistic
equations. The second study  focused on abstracts text information retrieval where each electronic
Parameter Base Value Influence
Database scale 1.0 -359.6
Fraction of query words which are striped 0.0 278.4
Disk bandwidth (Mbit/sec) 10.4 112.7
Compression ratio 0.5 -67.4
Multiprogramming level (per host) 4 -48.1
CPU speed (MIPS) 20.0 47.7
Posting size (bits) 40.0 -44.5
Hosts 1 -27.9
Disks per I/O bus 4 25.4
I/O bus bandwidth (Mbit/sec) 24.0 11.2
Buffer overhead (ms) 4.0 -9.33
Disk buffer size (Kbyte) 32 9.12
LAN bandwidth (Mbit/s) (4 hosts) 100.0 2.33
I/O bus overhead (ms) 0.0 -1.96
Disk seek time (ms) 6.0 -1.93
Bytes per block 512 -0.81
Instructions per byte for a merge 40 0.0
Answer entry size (bytes) 4.0 0.0
Instructions per byte of decompression 40 0.0
Instruction count per query 500,000 0.0
Cache size (postings) 0 0.0
Instructions per byte of union operation 5 0.0
Subquery instruction count 100,000 0.0
Instructions per disk fetch 10,000 0.0
LAN overhead (ms) 0.1 0.0
LAN bandwidth (Mb/s) 100.0 0.0
Subquery length (bytes) 1024 0.0
Table 3: A ranking of the influence of simulation parameters on response time for the systemindex organization
with Prefetch I query optimization.
abstract is an abstract of a paper document. In this form of retrieval, the inverted index records only the occurrence
of a word in an abstract, and not every occurrence. This dramatically reduces the size of the index with
respect to full-text retrieval.
In general, our results indicate that the host index organization is a good choice, especially if long inverted
lists are striped across disks. Long inverted lists are present in full-text information retrieval. Since the lists are
long, the bottleneck is I/O performance. The host index organization uses system resources effectively and can
lead to high query throughputs in many cases. When it does not perform the best, it is close to the best strategy.
For an application where only abstracts are indexed, the system organization (with the Prefetch I optimization)
actually outperforms the host organization. The bottleneck for these systems is the network. This is because
the inverted lists are much shorter, and can be easily moved across machines.
To study the impact of the experimental parameters on response time, we focus on the second study. Our
inverted list model and query model were based on inverted lists of actual abstracts and traces of actual user
queries from the Stanford University FOLIO information retrieval system. In both studies, query processing and
hardware measurement where accomplished by using a sophisticated simulation containing over 28 parameters.
Table 3 lists the parameters and the default values of each parameter. For each parameter in the table, a simulation
experiment was run which linearly varied the values of the parameter. The simulation reflects the the architecture
shown in Figure 1, as determined by the number of hosts, I/O buses and disks shown in the table. Full details of
our experiments and our results are available in the references.
One way to succinctly show the parameters involved in the studies and their influence on performance is to
“rank” themby their (normalized) influence. Here we only look at query response time as the performancemetric.
In particular, if a and b are the smallest and largest values measured for a parameter and x is the response time
for a and y the response time for b, we compute (y x)=(a=b) as an estimate of the influence the parameter has
on response time. Of course, this measure is only a rough indication of influence. The measure depends on the
ranges of values over which a parameter is measured. It also assumes that response time is monotonic over the
range of values chosen. We have inspected the data to insure that this last condition holds.
Table 3 shows the ranking of 28 parameters for the systemindex organization, as described in Section 4, with
the Prefetch I query optimization, as described in Section 5. In previous work, the systemindex organizationwas
shown to be the best overall choice for an index organization for abstracts text information retrieval. The positive
or negative nature of the ranking is due to the positive or negative influence the parameter has on response time.
Database scale has the strongest influence – this parameter linearly scales the length of an inverted list and
scales the lengths of all other objects in the system – such as the size of the answers to queries. With striping, a
fraction of the inverted lists (in particular the longest ones) are striped across the diskswithin a computer system.
This is a complementary technique to the list partitioningdone by the basic index organizationwe have discussed,
and can be very beneficial. Disk bandwidth is important due to the disk intensive nature of the computation.
The compression ratio linearly scales the length of the inverted lists, but does scale any other parameter. The
multiprogramming level is the number of simultaneous jobswhich are run on each host. The relative CPU speed
scales all computations which compute the number of instructions needed to accomplish a task. The posting size
is the number of bits needed to represent a posting. Hosts represents the number of processors in the system.
When this parameter is increased, a copy of the processor is made. That is, if the parameter doubles, the number
of I/O buses and disks in the entire system also doubles. In addition, the workload doubles, since the number of
concurrent queries is allocated on a per host basis. Examining the parameters at the end of the table, we see that
within the accuracy of the measurement, several parameters have no influence on response time. One surprising
fact shows cache size as having no influence. In fact, caches have no influence on response time, but have a
tremendous influence on throughput. Essentially, each query almost always has a cache miss. Thus, the response
time of the query is dictated by the read from disk of the cache miss and thus the cache has little influence on
response time. However, most queries have cache hits also, which dramatically improves throughput.
In this article, we have sampled issues in parallel information retrieval. As an introduction to the issues involved,
we have discussed the literature in the area to introduce the various areas of research. We then focused on a
specific example to illustrate the issued involved in distributed shared-nothing information retrieval, and discuss
physical index organization and query optimization techniques. Then, to give the reader a sense of the important
variables in the design of a system, we ranked the various parameters in an experimental simulation study in terms
of their influence on the response time of query processing.
 Ijsbrand Jan Aalbersberg and Frans Sijstermans. High-quality and high-performance full-text document retrieval: the
parallel infoguide system. In Proceedings of the First International Conference on Parallel and Distributed Information
Systems, pages 151–158, Miami Beach, Florida, 1991.
 Forbes J. Burkowski. Retrieval performance of a distributed text database utilizing a parallel processor document
server. In Proceedings of the Second International Symposium on Databases in Parallel and Distributed Systems,
pages 71–79, Dublin, Ireland, 1990.
 Janey K. Cringean, Roger England, Gordon A. Manson, and Peter Willett. Parallel text searching in serial files using
a processor farm. In Proceedings of Special Interest Group on Information Retrieval (SIGIR), pages 429–453, 1990.
 Samuel DeFazio. Full-text document retrieval benchmark. In Jim Gray, editor, The Benchmark Handbook for
Database and Transaction Processing Systems, chapter 8. Morgan Kaufmann, second edition, 1993.
 Samuel DeFazio and Joe Hull. Toward servicing textual database transactions on symmetric shared memory multiprocessors.
Proceedings of the Int’lWorkshop on High Performance Transaction Systems, Asilomar, 1991.
 Perry Alan Emrath. Page Indexing for Textual Information Retrieval Systems. PhD thesis, University of Illinois at
Urbana-Champaign, October 1983.
 Christos Faloutsos. Access methods for text. ACM Computing Surveys, 17:50–74, 1985.
 J. Fedorowicz. Database performance evaluation in an indexed file environment. ACM Transactions on Database
Systems, 12(1):85–110, 1987.
 William B. Frakes and Ricardo Baeza-Yates. Information Retrieval: Data Structures and Algorithms. Prentice-Hall,
 Gaston H. Gonnet, Ricardo A. Baeza-Yates, and Tim Snider. Lexicographical indices for text: Inverted files vs. PAT
trees. Technical Report OED-91-01, University ofWaterloo Centre for the New Oxford English Dictionary and Text
Research, Canada, 1991.
 Lee A.Hollaar. Implementations and evaluation of a parallel text searcher for very large text databases. In Proceedings
of the Twenty-Fifth Hawaii International Conference on System Sciences, pages 300–307. IEEE Computer society
 Byeong-Soo Jeong and Edward Omiecinski. Inverted file partitioning schemes for a shared-everythingmultiprocessor.
Technical Report GIT-CC-92/39, Georgia Institute of Technology, College of Computing, 1992.
 Donald E. Knuth. The Art of Computer Programming. Addison-Wesley, Reading, Massachusetts, 1973.
 Zheng Lin. Cat: An execution model for concurrent full text search. In Proceedings of the First International Conference
on Parallel and Distributed Information Systems, pages 151–158, Miami Beach, Florida, 1991.
 PatrickMartin, Ian A. Macleod, and Brent Nordin. A design of a distributed full text retrieval system. In Proceedings
of Special Interest Group on Information Retrieval (SIGIR), pages 131–137, Pisa, Italy, September 1986.
 Gerard Salton. Automatic Text Processing. Addison-Wesley, New York, 1989.
 Bruce Raymond Schatz. Interactive retrieval in information spaces distributed across a wide-area network. Technical
Report 90-35, University of Arizona, December 1990.
 Kurt Shoens, Anthony Tomasic, and Hector Garcia-Molina. Syntheticworkload performance analysis of incremental
updates. In Proceedings of Special Interest Group on Information Retrieval (SIGIR), Dublin, Ireland, 1994.
 Craig Stanfill. Partitioned posting files: A parallel inverted file structure for information retrieval. In Proceedings of
Special Interest Group on Information Retrieval (SIGIR), 1990.
 Craig Stanfill and Brewster Kahle. Parallel free-text search on the connection machine system. Communications of
the ACM, 29:1229–1239, 1986.
 Craig Stanfill, Robert Thau, and DavidWaltz. A parallel indexed algorithmfor information retrieval. In Proceedings
of the Twelfth Annual International ACM/SIGIR Conference on Research and Development in Information Retrieval,
pages 88–97, Cambridge, Massachusetts, 1989.
 Harold S. Stone. Parallel querying of large databases: A case study. IEEE Computer, pages 11–21, October 1987.
 Anthony Tomasic and Hector Garcia-Molina. Caching and database scaling in distributed shared-nothing information
retrieval systems. In Proceedings of the Special Interest Group onManagement ofData (SIGMOD),Washington,D.C.,
 Anthony Tomasic and Hector Garcia-Molina. Performance of inverted indices in shared-nothing distributed text document
information retrieval systems. In Proceedings of the Second International Conference On Parallel and Distributed
Information Systems, San Diego, 1993.
 Anthony Tomasic and Hector Garcia-Molina. Query processing and inverted indices in shared-nothing document
information retrieval systems. The VLDB Journal, 2(3):243–271, July 1993.
 Anthony Tomasic, Hector Garcia-Molina, and Kurt Shoens. Incremental updates of inverted lists for text
document retrieval. Technical Note STAN-CS-TN-93-1, Stanford University, 1993. Available via FTP
 Anthony Tomasic, Hector Garcia-Molina, and Kurt Shoens. Incremental updates of inverted lists for text document
retrieval. In Proceedings of 1994 ACM SIGMOD International Conference on Management of Data, Minneapolis,
 Howard R. Turtle and W. Bruce Croft. Uncertainty in information retrieval systems. In Amihai Motro and Philippe
Smets, editors, Proceedings of the Workshop on Uncertainty Management in Information Systems, pages 111–137,
Mallorca, Spain, September 1992.
 Peter Weiss. Size Reduction of Inverted Files Using Data Compression and Data Structure Reorganization. PhD
thesis, George Washington University, 1990.
 Justin Zobel, Alistair Moffat, and Ron Sacks-Davis. An efficient indexing technique for full-text database systems.
In Proceedings of 18th International Conference on Very Large Databases, Vancouver, 1992.
IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING
CALL FOR PAPERS
Research Surveys and Correspondences on Recent Developments
We are interested to publish in the IEEE Transactions on Knowledge and Data
Engineering research surveys and correspondences on recent developments. These
two types of articles are found to have greater influence in the work of the majority
of our readers.
Research surveys are articles that present new taxonomies, research issues,
and current directions on a specific topic in the knowledge and data engineering
areas. Each article should have an extensive bibliography that is useful for experts
working in the area and should not be tutorial in nature. Correspondences on
recent developments are articles that describe recent results, prototypes, and new
Submissions will be reviewed using the same standard as other regular submissions.
Since these articles have greater appeal to our readers, we will publish
these articles in the next available issue once they are accepted.
Address to send articles: Benjamin W. Wah, Editor-in-Chief
Coordinated Science Laboratory
University of Illinois, Urbana-Champaign
1308 West Main Street
Urbana, IL 61801, USA
Phone: (217) 333-3516 (office), 244-7175 (sec./fax)
Submission Deadline: None
Reviewing Delay: One month for correspondences, three months for surveys
Publication Delay: None; articles are published as soon as they are accepted
Submission Guidelines: See the inside back cover of any issue of TKDE or by
anonymous ftp from manip.crhc.uiuc.edu (126.96.36.199)
in file /pub/tkde/submission.guide.ascii
Length Requirements: 40 double-spaced pages for surveys, 6 double-spaced pages
Areas of Interest: See the editorial in the February’94 issue of TKDE or by
anonymous ftp from manip.crhc.uiuc.edu in file
IEEE Computer Society
1730Massachusetts Ave, NW
Washington, D.C. 20036-1903
Silver Spring, MD