Chapter 20: Parallel Databases

 Introduction
 I/O Parallelism
 Interquery Parallelism
 Intraquery Parallelism
 Intraoperation Parallelism
 Interoperation Parallelism
 Design of Parallel Systems

20.1
 Introduction

 Parallel machines are becoming quite common and affordable

 Prices of microprocessors, memory and disks have dropped sharply
 Databases are growing increasingly large
 large volumes of transaction data are collected and stored for later
analysis.
 multimedia objects like images are increasingly stored in databases
 Large-scale parallel database systems increasingly used for:
 storing large volumes of data
 processing time-consuming decision-support queries
 providing high throughput for transaction processing

20.2
 Parallelism in Databases

 Data can be partitioned across multiple disks for parallel I/O.

 Individual relational operations (e.g., sort, join, aggregation) can
be executed in parallel
 data can be partitioned and each processor can work independently
on its own partition.
 Queries are expressed in high level language (SQL, translated to
relational algebra)
 makes parallelization easier.
 Different queries can be run in parallel with each other.
Concurrency control takes care of conflicts.
 Thus, databases naturally lend themselves to parallelism.

20.3
 I/O Parallelism

 Reduce the time required to retrieve relations from disk by partitioning

the relations on multiple disks.
 Horizontal partitioning – tuples of a relation are divided among many
disks such that each tuple resides on one disk.
 Partitioning techniques (number of disks = n):
Round-robin:
Send the ith tuple inserted in the relation to disk i mod n.
 The scheme is ideally suited for applications that wish to read the
entire relation sequentially for each query.
Hash partitioning:
 Choose one or more attributes as the partitioning attributes.
 Choose hash function h with range 0…n - 1
 Let i denote result of hash function h applied to the partitioning attribute
value of a tuple. Send tuple to disk i.

20.4
 I/O Parallelism (Cont.)

 For example, if a relation is partitioned on the telephone number

attribute, then we can answer the query “Find the record of the
employee with telephone-number = 555-3333” by applying the
partitioning hash function to 555-3333 and then searching that
disk.
 Directing a query to a single disk saves the startup cost of
initiating a query on multiple disks, and leaves the other disks
free to process other queries.

 Range partitioning:
 Choose an attribute as the partitioning attribute.
 A partitioning vector [vo, v1, ..., vn-2] is chosen.
 Let v be the partitioning attribute value of a tuple. Tuples such that vi  vi+1 go to disk I +
1. Tuples with v < v0 go to disk 0 and tuples with v  vn-2 go to disk n-1.
 E.g., For example, range partitioning with three disks numbered 0, 1, and 2 may assign
tuples with values less than 5 to disk 0, values between 5 and 40 to disk1, and values
greater than 40 to disk 2.

20.5
Comparison of Partitioning Techniques

 Evaluate how well partitioning techniques support the following

types of data access:
1.Scanning the entire relation.
2.Locating a tuple associatively – point queries.
 E.g., r.A = 25.
3.Locating all tuples such that the value of a given attribute lies
within a specified range – range queries.
 E.g., 10  r.A < 25.

20.6
Comparison of Partitioning Techniques (Cont.)

Round robin:
 Advantages
 Best suited for sequential scan of entire relation on each query.
 All disks have almost an equal number of tuples; retrieval work is
thus well balanced between disks.
 Range queries are difficult to process
 No clustering -- tuples are scattered across all disks

20.7
Comparison of Partitioning Techniques(Cont.)

Hash partitioning:
 Good for sequential access
 Assuming hash function is good, and partitioning attributes form a
key, tuples will be equally distributed between disks
 Retrieval work is then well balanced between disks.
 Good for point queries on partitioning attribute
 Can lookup single disk, leaving others available for answering other
queries.
 Index on partitioning attribute can be local to disk, making lookup
and update more efficient
 No clustering, so difficult to answer range queries

20.8
Comparison of Partitioning Techniques (Cont.)

Range partitioning:
 Provides data clustering by partitioning attribute value.
 Good for sequential access
 Good for point queries on partitioning attribute: only one disk
needs to be accessed.
 For range queries on partitioning attribute, one to a few disks
may need to be accessed
 Remaining disks are available for other queries.
 Good if result tuples are from one to a few blocks.
 If many blocks are to be fetched, they are still fetched from one
to a few disks, and potential parallelism in disk access is wasted

20.9
 Partitioning a Relation across Disks

 If a relation contains only a few tuples which will fit into a single
disk block, then assign the relation to a single disk.
 Large relations are preferably partitioned across all the available
disks.
 If a relation consists of m disk blocks and there are n disks
available in the system, then the relation should be allocated
min(m,n) disks.

20.10
 Handling of Skew

 The distribution of tuples to disks may be skewed — that is, some disks
have many tuples, while others may have fewer tuples.
 Types of skew:
 Attribute-value skew.
 Some values appear in the partitioning attributes of many tuples; all the
tuples with the same value for the partitioning attribute end up in the
same partition.
 Can occur with range-partitioning and hash-partitioning.
 For example, 1000 tuples divided on 10 disk then 100

tuples on each disk.

 Partition skew.
 With range-partitioning, badly chosen partition vector may assign too
many tuples to some partitions and too few to others.
 Less likely with hash-partitioning if a good hash-function is chosen.

20.11
 Handling Skew in Range-Partitioning

 To create a balanced partitioning vector (assuming partitioning

attribute forms a key of the relation):
 Sort the relation on the partitioning attribute.
 Construct the partition vector by scanning the relation in sorted order
as follows.
 After every 1/nth of the relation has been read, the value of the
partitioning attribute of the next tuple is added to the partition
vector.
 n denotes the number of partitions to be constructed.
 Duplicate entries or imbalances can result if duplicates are present in
partitioning attributes.
 Alternative technique based on histograms used in practice

20.12
 Handling Skew using Histograms
 Balanced partitioning vector can be constructed from histogram in a
relatively straightforward fashion
 Assume uniform distribution within each range of the histogram
 Histogram can be constructed by scanning relation, or sampling
(blocks containing) tuples of the relation

20.13
Handling Skew Using Virtual Processor
Partitioning

 Skew in range partitioning can be handled elegantly using

virtual processor partitioning:
 create a large number of partitions (say 10 to 20 times the number
of processors)
 Assign virtual processors to partitions either in round-robin fashion
or based on estimated cost of processing each virtual partition
 Basic idea:
 If any normal partition would have been skewed, it is very likely the
skew is spread over a number of virtual partitions
 Skewed virtual partitions get spread across a number of processors,
so work gets distributed evenly!

20.14
 Interquery Parallelism
 Queries/transactions execute in parallel with one another.
 Increases transaction throughput; used primarily to scale up a
transaction processing system to support a larger number of
transactions per second.
 Easiest form of parallelism to support, particularly in a shared-
memory parallel database, because even sequential database
systems support concurrent processing.
 More complicated to implement on shared-disk or shared-
nothing architectures
 Locking and logging must be coordinated by passing messages
between processors.
 Data in a local buffer may have been updated at another processor.
 Cache-coherency has to be maintained — reads and writes of data
in buffer must find latest version of data.

20.15
 Cache Coherency Protocol

 Example of a cache coherency protocol for shared disk systems:

 Before reading/writing to a page, the page must be locked in
shared/exclusive mode.
 On locking a page, the page must be read from disk
 Before unlocking a page, the page must be written to disk if it was
modified.
 More complex protocols with fewer disk reads/writes exist.
 Cache coherency protocols for shared-nothing systems are
similar. Each database page is assigned a home processor.
Requests to fetch the page or write it to disk are sent to the
home processor.

20.16
 Intraquery Parallelism

 Execution of a single query in parallel on multiple

processors/disks; important for speeding up long-running
queries.
 Two complementary forms of intraquery parallelism :
 Intraoperation Parallelism – parallelize the execution of each
individual operation in the query.
 Interoperation Parallelism – execute the different operations in a
query expression in parallel.
the first form scales better with increasing parallelism because
the number of tuples processed by each operation is typically
more than the number of operations in a query

20.17
Parallel Processing of Relational Operations

 Our discussion of parallel algorithms assumes:

 read-only queries
 shared-nothing architecture
 n processors, P0, ..., Pn-1, and n disks D0, ..., Dn-1, where disk Di is
associated with processor Pi.
 If a processor has multiple disks they can simply simulate a
single disk Di.
 Shared-nothing architectures can be efficiently simulated on
shared-memory and shared-disk systems.
 Algorithms for shared-nothing systems can thus be run on shared-
memory and shared-disk systems.
 However, some optimizations may be possible.

20.18
 Parallel Sort
Range-Partitioning Sort
 Choose processors P0, ..., Pm, where m  n -1 to do sorting.
 Create range-partition vector with m entries, on the sorting attributes
 Redistribute the relation using range partitioning
 all tuples that lie in the ith range are sent to processor Pi
 Pi stores the tuples it received temporarily on disk Di.
 This step requires I/O and communication overhead.
 Each processor Pi sorts its partition of the relation locally.
 Each processors executes same operation (sort) in parallel with other
processors, without any interaction with the others (data parallelism).
 Final merge operation is trivial: range-partitioning ensures that, for 1 j
m, the key values in processor Pi are all less than the key values in Pj.

20.19
 Parallel Sort (Cont.)

Parallel External Sort-Merge

 Assume the relation has already been partitioned among disks
D0, ..., Dn-1 (in whatever manner).
 Each processor Pi locally sorts the data on disk Di.
 The sorted runs on each processor are then merged to get the
final sorted output.
 Parallelize the merging of sorted runs as follows:
 The sorted partitions at each processor Pi are range-partitioned
across the processors P0, ..., Pm-1.
 Each processor Pi performs a merge on the streams as they are
received, to get a single sorted run.
 The sorted runs on processors P0,..., Pm-1 are concatenated to get
the final result.

20.20
 Parallel Join

 The join operation requires pairs of tuples to be tested to see if

they satisfy the join condition, and if they do, the pair is added to
the join output.
 Parallel join algorithms attempt to split the pairs to be tested over
several processors. Each processor then computes part of the
join locally.
 In a final step, the results from each processor can be collected
together to produce the final result.

20.21
 Partitioned Join

 For equi-joins and natural joins, it is possible to partition the two input
relations across the processors, and compute the join locally at each
processor.
 Let r and s be the input relations, and we want to compute r r.A=s.B s.
 r and s each are partitioned into n partitions, denoted r0, r1, ..., rn-1 and
s0, s1, ..., sn-1.
 Can use either range partitioning or hash partitioning.
 r and s must be partitioned on their join attributes r.A and s.B), using
the same range-partitioning vector or hash function.
 Partitions ri and si are sent to processor Pi,
 Each processor Pi locally computes ri ri.A=si.B si. Any of the standard
join methods can be used.

20.22
Partitioned Join (Cont.)

20.23
 Fragment-and-Replicate Join

 Partitioning not possible for some join conditions

 e.g., non-equijoin conditions, such as r.A > s.B.
 For joins were partitioning is not applicable, parallelization can
be accomplished by fragment and replicate technique
 Depicted on next slide
 Special case – asymmetric fragment-and-replicate:
 One of the relations, say r, is partitioned; any partitioning technique
can be used.
 The other relation, s, is replicated across all the processors.
 Processor Pi then locally computes the join of ri with all of s using
any join technique.

20.24
Depiction of Fragment-and-Replicate Joins

a. Asymmetric
Fragment and
Replicate b. Fragment and Replicate
20.25
 Fragment-and-Replicate Join (Cont.)

 General case: reduces the sizes of the relations at each

processor.
 r is partitioned into n partitions,r0, r1, ..., r n-1;s is partitioned into m
partitions, s0, s1, ..., sm-1.
 Any partitioning technique may be used.
 There must be at least m * n processors.
 Label the processors as
 P0,0, P0,1, ..., P0,m-1, P1,0, ..., Pn-1m-1.
 Pi,j computes the join of ri with sj. In order to do so, ri is replicated to
Pi,0, Pi,1, ..., Pi,m-1, while si is replicated to P0,i, P1,i, ..., Pn-1,i
 Any join technique can be used at each processor Pi,j.

20.26
 Fragment-and-Replicate Join (Cont.)

 Both versions of fragment-and-replicate work with any join

condition, since every tuple in r can be tested with every tuple in s.
 Usually has a higher cost than partitioning, since one of the
relations (for asymmetric fragment-and-replicate) or both relations
(for general fragment-and-replicate) have to be replicated.
 Sometimes asymmetric fragment-and-replicate is preferable even
though partitioning could be used.
 E.g., say s is small and r is large, and already partitioned. It may be
cheaper to replicate s across all processors, rather than repartition r
and s on the join attributes.

20.27
 Partitioned Parallel Hash-Join

Parallelizing partitioned hash join:

 Assume s is smaller than r and therefore s is chosen as the build
relation.
 A hash function h1 takes the join attribute value of each tuple in s
and maps this tuple to one of the n processors.
 Each processor Pi reads the tuples of s that are on its disk Di,
and sends each tuple to the appropriate processor based on
hash function h1. Let si denote the tuples of relation s that are
sent to processor Pi.
 As tuples of relation s are received at the destination processors,
they are partitioned further using another hash function, h2,
which is used to compute the hash-join locally. (Cont.)

20.28
 Partitioned Parallel Hash-Join (Cont.)

 Once the tuples of s have been distributed, the larger relation r is

redistributed across the m processors using the hash function h1
 Let ri denote the tuples of relation r that are sent to processor Pi.
 As the r tuples are received at the destination processors, they
are repartitioned using the function h2
 (just as the probe relation is partitioned in the sequential hash-join
algorithm).
 Each processor Pi executes the build and probe phases of the
hash-join algorithm on the local partitions ri and s of r and s to
produce a partition of the final result of the hash-join.
 Note: Hash-join optimizations can be applied to the parallel case
 e.g., the hybrid hash-join algorithm can be used to cache some of
the incoming tuples in memory and avoid the cost of writing them and
reading them back in.

20.29
 Parallel Nested-Loop Join

 Assume that
 relation s is much smaller than relation r and that r is stored by
partitioning.
 there is an index on a join attribute of relation r at each of the
partitions of relation r.
 Use asymmetric fragment-and-replicate, with relation s being
replicated, and using the existing partitioning of relation r.
 Each processor Pj where a partition of relation s is stored reads
the tuples of relation s stored in Dj, and replicates the tuples to
every other processor Pi.
 At the end of this phase, relation s is replicated at all sites that store
tuples of relation r.
 Each processor Pi performs an indexed nested-loop join of
relation s with the ith partition of relation r.

20.30
 Other Relational Operations

Selection (r)
 If  is of the form ai = v, where ai is an attribute and v a value.
 If r is partitioned on ai the selection is performed at a single
processor.
 If  is of the form l <= ai <= u (i.e.,  is a range selection) and the
relation has been range-partitioned on ai
 Selection is performed at each processor whose partition overlaps
with the specified range of values.
 In all other cases: the selection is performed in parallel at all the
processors.

20.31
 Other Relational Operations (Cont.)

 Duplicate elimination
 Perform by using either of the parallel sort techniques
 eliminate duplicates as soon as they are found during sorting.
 Can also partition the tuples (using either range- or hash-
partitioning) and perform duplicate elimination locally at each
processor.

 Projection
 Projection without duplicate elimination can be performed as tuples
are read in from disk in parallel.
 If duplicate elimination is required, any of the above duplicate
elimination techniques can be used.

20.32
 Grouping/Aggregation

 Partition the relation on the grouping attributes and then

compute the aggregate values locally at each processor.
 Can reduce cost of transferring tuples during partitioning by
partly computing aggregate values before partitioning.
 Consider the sum aggregation operation:
 Perform aggregation operation at each processor Pi on those tuples
stored on disk Di
 results in tuples with partial sums at each processor.
 Result of the local aggregation is partitioned on the grouping
attributes, and the aggregation performed again at each processor
Pi to get the final result.
 Fewer tuples need to be sent to other processors during
partitioning.

20.33
Cost of Parallel Evaluation of Operations

 If there is no skew in the partitioning, and there is no overhead

due to the parallel evaluation, expected speed-up will be 1/n
 If skew and overheads are also to be taken into account, the
time taken by a parallel operation can be estimated as
Tpart + Tasm + max (T0, T1, …, Tn-1)
 Tpart is the time for partitioning the relations
 Tasm is the time for assembling the results
 Ti is the time taken for the operation at processor Pi
 this needs to be estimated taking into account the skew, and the
time wasted in contentions.

20.34
 Interoperator Parallelism

 Pipelined parallelism
 Consider a join of four relations
 r1 r2 r3 r4
 Set up a pipeline that computes the three joins in parallel
 Let P1 be assigned the computation of
temp1 = r1 r2
 And P2 be assigned the computation of temp2 = temp1 r3
 And P3 be assigned the computation of temp2 r4
 Each of these operations can execute in parallel, sending result
tuples it computes to the next operation even as it is computing
further results
 Provided a pipelineable join evaluation algorithm (e.g. indexed
nested loops join) is used

20.35
 Factors Limiting Utility of Pipeline
Parallelism

 Pipeline parallelism is useful since it avoids writing intermediate

results to disk
 Useful with small number of processors, but does not scale up
well with more processors. One reason is that pipeline chains do
not attain sufficient length.
 Cannot pipeline operators which do not produce output until all
inputs have been accessed (e.g. aggregate and sort) 
 Little speedup is obtained for the frequent cases of skew in which
one operator's execution cost is much higher than the others.

20.36
 Independent Parallelism

 Independent parallelism
 Consider a join of four relations
r1 r2 r3 r4
 Let P1 be assigned the computation of
temp1 = r1 r2
 And P2 be assigned the computation of temp2 = r3 r4
 And P3 be assigned the computation of temp1 temp2
 P1 and P2 can work independently in parallel
 P3 has to wait for input from P1 and P2
– Can pipeline output of P1 and P2 to P3, combining
independent parallelism and pipelined parallelism
 Does not provide a high degree of parallelism
 useful with a lower degree of parallelism.
 less useful in a highly parallel system,

20.37
 Query Optimization

 Query optimization in parallel databases is significantly more complex

than query optimization in sequential databases.
 Cost models are more complicated, since we must take into account
partitioning costs and issues such as skew and resource contention.
 When scheduling execution tree in parallel system, must decide:
 How to parallelize each operation and how many processors to use for it.
 What operations to pipeline, what operations to execute independently in
parallel, and what operations to execute sequentially, one after the other.
 Determining the amount of resources to allocate for each operation is a
problem.
 E.g., allocating more processors than optimal can result in high
communication overhead.
 Long pipelines should be avoided as the final operation may wait a lot
for inputs, while holding precious resources

20.38
 Query Optimization (Cont.)
 The number of parallel evaluation plans from which to choose from is much
larger than the number of sequential evaluation plans.
 Therefore heuristics are needed while optimization
 Two alternative heuristics for choosing parallel plans:
 No pipelining and inter-operation pipelining; just parallelize every operation
across all processors.
 Finding best plan is now much easier --- use standard optimization
technique, but with new cost model
 Volcano parallel database popularize the exchange-operator model
– exchange operator is introduced into query plans to partition and
distribute tuples
– each operation works independently on local data on each processor, in
parallel with other copies of the operation
 First choose most efficient sequential plan and then choose how best to
parallelize the operations in that plan.
 Can explore pipelined parallelism as an option
 Choosing a good physical organization (partitioning technique) is important
to speed up queries.

20.39
 Design of Parallel Systems

Some issues in the design of parallel systems:

 Parallel loading of data from external sources is needed in order
to handle large volumes of incoming data.
 Resilience to failure of some processors or disks.
 Probability of some disk or processor failing is higher in a parallel
system.
 Operation (perhaps with degraded performance) should be possible
in spite of failure.
 Redundancy achieved by storing extra copy of every data item at
another processor.

20.40
 Design of Parallel Systems (Cont.)

 On-line reorganization of data and schema changes must be

supported.
 For example, index construction on terabyte databases can take
hours or days even on a parallel system.
 Need to allow other processing (insertions/deletions/updates) to
be performed on relation even as index is being constructed.
 Basic idea: index construction tracks changes and ``catches up'‘ on
changes at the end.
 Also need support for on-line repartitioning and schema changes
(executed concurrently with other processing).

20.41
End of Chapter