Optimized Access Strategies for a Distributed Database Design

Rajinder Singh, Gurvinder Singh & Varinder Pannu
International Journal of Data Engineering (IJDE), Volume (2) : Issue (3) : 2011 102
Optimized Access Strategies for a Distributed Database Design
Rajinder Singh tovirk@yahoo.com
Assoc. Professor
Faculty of Computer Science & Engineering
Guru Nanak Dev University, Amritsar-143001
Punjab, India.
Gurvinder Singh gsbawa71@yahoo.com
Assoc. Professor
Guru Nanak Dev University, Amritsar-143001
Punjab, India.
Varinder Pannu viki_virk@yahoo.com
Computer Engineer
Govt. Polytechnic, Amritsar-143001
Punjab, India.
Abstract
Distributed Database Query Optimization is achieved thru many complex sub operations on the
Relations, Network Sites, Local Processing Facilities and the Database System itself. Many of
these sub problems are NP-Hard itself, which makes Distributed Database Query Optimization a
very complex and hard process. One of these NP Hard components is optimal allocation of
various sub-queries to different sites of data distribution. Most of prevalent solutions take help of
Exhaustive Enumeration Techniques, along with use of innovative heuristics. In this Paper we
have proposed a stochastic model simulating a Distributed Database environment, and shown
benefits of using innovative Genetic Algorithms (GA) for optimizing the sequence of sub-query
operations allocation over the Network Sites. Also, the effect of varying Genetic Parameters on
Solution’s quality is analyzed.
Keywords: Distributed Query Optimization, Database Statistics, Query Execution Plan, Genetic
Algorithms, Operation Allocation.
1. INTRODUCTION
Query Optimization process involves finding a near optimal query execution plan which
represents the overall execution strategy for the query. The efficiency of distributed database
system is significantly dependent on the extent of optimality of this execution plan. According to
Ozsu and Valduriez[1],this process of generating a good query execution strategy involves three
phases. First is to find a search space which is a set of alternative execution plans for query.
Second is to build a cost model which can compare costs of different execution plans. Finally in
third step we explore a search strategy to find the best possible execution plan using cost model.
Before putting any queries to a Distributed Database, one needs to design it according to the
needs of an organization. Analysts have to plan data/Fragment allocation according to the nature
and frequencies of the various queries at different sites. Different Sequence of operations or sub-
operations needed to generate results of a query is very large e.g. as near to n! for relational join
of n relations. Decisions have to be taken dynamically for allocation of intermediate relation
fragments generated during the query. A Transaction Profile is build to provide this information for
various queries to the database. This paper gives a Genetic Algorithm for this step of Execution
Plan. It assumes that a transaction profile provides the necessary details of fragment allocation
and cost profiles for various pair of sites. It gives a cost model to predict cost of various allocation

plans for intermediate relations and sub operations generated during process of Query. Finally
this GA finds the best possible execution strategy in terms of finding sequence and sites for
various sub operations, called Operation Allocation Problem[2].
A GA(Genetic Algorithm) has several advantages over other approaches, as it has been
successfully applied to a vast set of real world applications, not only that it gives a robust
solution but also a good set of alternative possible solutions to choose from [3] and is inherently
parallel to achieve good response time.
2. PREVIOUS RESEARCH WORK
Distributed database systems design and query optimization has been and will remain an active
area of research for a lot times to come, due to complex and intractable nature of the
problem[4,5,6,7,8,9,10].Most of the work has concentrated on two aspects: Data Allocation(The
plan of allocating Fragments to various sites) and Operation Allocation(How to generate a
sequence of subqueries on various sites). Apers and P.M have discussed in detail the data
allocation problem and their fragmentation in [11]. An integrated solution to problems of Data
Fragmentation, allocation, replication in Distributed Databases, has been proposed in
Tamhankar & Ram[12]. Zehai Zhou[10] propose using heuristics and genetic algorithms for large
scale database query optimization.The NP Hard problem is reduced to a join ordering problem
similar to a variant of a Travelling Salesman Problem.Several heuristics and a GA is proposed for
solving the join order problem.
Simulation experiments for comparison of Branch & Bound, Simulated Annealing, Greedy
approaches for operation allocation problem have been describes in detail by Martin & Lam[13].
Frieder and Baru [14] propose dynamic site selection strategies for distributed database design
on a microcomputer. March & Rho in [2] have proposed an excellent cost model for reducing local
I/O costs, CPU Costs and Communication Costs in operation allocation strategy. Johansonn &
Noumann in [15] extended their work bu considering parallel processing and Load Balancing in
Data and Operation Allocation.
3. Objective Function & Cost Model
Database Statistics
The main factor affecting the performance of an execution strategy is the size of the intermediate
fragments produced during the execution of the sub operations of the query. As many
intermediate relations or fragments will need to move over various sites, we need to estimate the
size of them to determine transmission costs. This estimation is based on statistical information
about base relation and formulas to predict the cardinalities of the results of operations [1].
The set of operations (sub-queries) generated in response to a query can be represented by an
operator tree. Nodes of operator tree represent various operations and lines represent cost
(based on size of fragment) of operation sequence. A site’s Local CPU and I/O costs are
proportional to the size in bytes (blocks) of data processed and communication costs depend on
communication coefficients between a pair of sites and bytes of blocks moved.
The main assumptions are that Transaction Profiles are known a-priori, providing the details of
frequencies of transaction at various sites, base relation sizes and allocation plan at various sites,
communication coefficients giving cost of communication amongst various pair of sites and local
I/O and CPU coefficients. Also query execution order is given, we emphasize on finding sub-
query allocation and cost associated to it with respect to allocation to various sites.
Projection, Selection and Joins account for most of the sub-queries in a Database Query and for
simplicity purposes, only these operations have been considered.

3.1 Objective Function Formulation
We start by simulating a design of distributed database by taking a set ‘S’ of data distribution
sites. Set ‘R’ of relations/fragments stored on those sites. A Set ‘Q’ as set of transactions.
Let a query transaction ‘q’ for retrieval, be broken into a set of ‘j’ sub queries on the ‘R’ set of
relations.
3.1.1 Decision Variables :-
(i) Data Allocation Variable Ars
Ars = 1 ( if site ‘S’ holds copy of relation/fragment ‘r’)
Ars = 0 (otherwise i.e. fragment ‘r’ copy is not available at set S)
(ii) Variables for site selection for sub query execution:
Sys
q
: ( Represents sequence of sub query execution at
various sites in the life time of query)
Sys
q
= 1 ( if subquery ‘y’ of Query ‘q’ is done at site s )
Sys
q
= 0 ( otherwise )
(iii) For Join operations a notation is proposed to handle left previous operation
operation of a join operation (LPO) & right previous operation of a join(RPO) as
following:
Syv[p]S = 1 ( for [p] = 1 for left previous operation of a Join )
Syv[p]S = 1 (for [p] = 2 for right previous operation of a Join )
Syv[p]S = 0 otherwise
(iv) I
q
ry represents the query tree in such a way that sub query ‘y’ of
query ‘q’ references the intermediate relation/fragment r.
I
q
ry = 1 ( if the base relation ‘r’ or intermediate fragment ‘r’ is
used by sub query ‘y’ of ‘q’ query)
I
q
ry = 0 otherwise
(v) For use of intermediate Relations by Join Operation
Iq
ryv[p] = 1 ( for lpo of join ‘y’ )
Iq
ryv[p] = 1 ( for rpo of join ‘y’)
Iq
ryv[p] = 0 otherwise
By making use of above decision variables operation allocation
problem formulation is represented as,
Given a input data file highlighting data allocation scheme matrix,
given by Sq
ys Data Allocation Scheme Matrix: A base relation y
stored at site s.
i.e.

Given a Transaction Profile,a Data Allocation Sceme represented by variable Ars
And given I
q
ry intermediate relations/fragments is used by sub query y of
query q
We have an objective Function to calculate as to find Sq
ys
3.1.2) Cost model
Given a set of fragments
R = {r1,r2,…,rn}
& a network of sites.
S = {s1,s2,…,sm}
& a set of sub queries
Q = {q1,q2,…,qq}
Sub Query Allocation problem involves finding the “optimal” possible distribution of R to S.
Ozsu gives a model for Total cost as Total Cost Function having two components: query
processing and storage cost as
TOC = ∑ QPCi + ∑vs€S∑ vfj€F STCjk
Where QPCi is query processing cost of application qi and STCjk is the cost of storing
fragment Fj at site Sk.
We choose Ozsu’s model of query cost as function of sum of local processing costs and
transmission costs .We simplify it further by ignoring update costs and ignoring concurrency
control costs as we are giving model for retrieval transactions(queries) only. Further concurrent
retrievals don’t impose any more integrity control costs.
Ozsu’s formulation gives
QPCi = PCi + TCi ( PC: Processing Cost, TC :Transmission Cost )
&
PCi = ACi + IEi + CCi ( AC: Access Costs, IE : Integrity Enforcement Costs,
CC : Concurrent Update control costs )
In our model we discard the sum of two costs components (IEi +CCi), because as discussed in
Para above, we present a simple model of retrieval queries only.
Therefore Access Costs may be represented as
ACi = ∑vs€S∑ vfj€F ( uij * URij + rij * RRij ) * xjk * LPCk
The summation gives total number of accesses for all the fragments referenced by qi. Now
xjk selects only those cost volume entries for sites where fragments are stored actually.

3.1.3) Local Processing Costs
For Simple selection & projections
LOPCq
y = ∑s Sq
ys(I POs ∑r Iq
ryMq
ry + CPCs ∑r Iq
ryMq
ry) (1)
Where M
q
ry = No. of memory blocks of relations ‘r’ accessed by sub-query y of q.
IPOs = Input Output Cost Coefficient of site s in msec per 8k bytes
CPCs = CPU Cost coefficient of site s.
So equation (1) represents local processing costs of transforming input relation from disk to
memory and CPU time for processing a Selection or Projection at sites s.
Ozsu’s model ignores join’s local processing cost details. For that we have extended this model
to add local join costs details as following.
Local processing costs for a join
LOPC
q
y = ∑sS
q
ysIPOs∑p∑rpvI
q
ryv[p]Mq 2(a)
+
∑sS
q
ys(IPOt∏rI
q
ryM
q
ry + CPCs∏rI
q
ryM
q
ry) 2(b)
Where pv is Selectivity Factor & is referred as the ratio of possible different values of a
field to the domain of that field.(0<= pv <=1)
Mryv[p] is the size of intermediate relation
where v[p] represents p=1 for left previous operation of a join &
p=2 for right previous operation of a join.
Equation 2(a) represents
Input Output costs in storing intermediate results of previous operations to the site of
current join operation.
Equation 2(b) represents
CPU & I/O costs for performing current join operations at site ‘s’.
3.1.4) Communication Costs:
An involved in case of join operations only as we have assured that selections & projections of
retrievals an to be done only at sites which hold a copy of that base relations. Join may be
performed at any of possible sites.
COMM
q
y = ∑p ∑s ∑t S
q
yv[m]s * S
q
ytCst ( ryv[p] Mq
ryv[p] )
Where
Cst ( is the communication cost coefficient taken from input data matrix )
Cst = 0 if (s = t) ( i.e. previous operations and join operation on same site )

If final operation is not done at the query destination site then a Communication component is
added separately for sending the final query result that site.
4. THE GENETIC ALGORITHM (GA_OA)
GA_OA ( Genetic Algorithm for Operation Allocation) starts by generating an initial pool of
solutions by random generation of operation sequence at given number of sites. An improvement
in this previous prevalent approach in [9][10] has been done in this GA is to use transaction
profile statistics to generate it.
Each chromosome is evaluated according to objective function and assigned a Fitness value
accordingly. Next populations are generated using principles of GA as in [3] i.e. applying
SELECTION,CROSSOVER & MUTATION, The fitter a member is more chance it gets to enter
the mating pool to generate next population.
Crossover is used so that off-spring shares features of both parents and possibly improves over
them. Mutation operator is applied with very small probability like.02 ,so as some important
features of parent population of chromosomes are not lost .Elitism is also applied, which ensure
that best chromosome of previous population enters next population by 1oo percent probability.
A sequence of integers like 2 1 3 3 4 1 is used to represent a chromosome such that it
represents the sub-query allocation plan in the way that, that sub-query 1 is done at site 2, sub-
query 2 is done at site1, sub-query 3 is done at site3, sub-query 4 is done at site 3, sub-query 5 is
done at site 4, sub-query 6 is done at site 1.
A Structured English representation of GA_OA is outlined below:
1. Generate an initial population pool of chromosomes, based on half the members of
the pool generated from transaction profiles, data allocation profiles and others
selecting randomly over no. of sites .
2. Evaluate the fitness of each member based on objective function to reduce the total
cost of a query.
3. Based on Stochastic remainder method select and give more chance to fitter
members to enter a mating pool according to probability proportional to their fitness
value.
4. To enable Elitism, enter the most fit member of previous generation in mating pool,
by replacing it with least fit.
5. Apply crossover (probability=0.7) and mutation (0.2) to generate a new population
pool.
6. Repeat steps 2 to 5 until maximum no of generations are generated.
5. EXPERIMENTS & RESULTS
Experiments were conducted after coding the GA_OA simulator on an Intel® core™ 2 6420 @
2.13 GHz machine with 1.00 GB RAM on WINDOWS-XP platform. Exhaustive Enumeration
simulator program was developed with varying all possible permutations of sub-query allocation
sequence. It was observed that Exhaustive Enumeration run time rises exponentially as
compared to GA when we increase no. of joins or no. of sites. When no. of joins are increased

from 4 to 10 Exhaustive_ Enumeration chokes very quickly but GA Run Time rises slowly. In case
of increasing the number of sites ,Run Time for GA increases linearly whereas Exhaustive_
Enumeration rises exponentially and very quickly becomes almost intractable.
The performance of simulator varied as we vary genetic operators Crossover and Mutation as
highlighted by the subsequent graphs, A crossover value of 0.6 was found giving optimal results
and mutation parameter of 0.2 achieved optimal solution in least number of generations.
200
250
45 sec
2Hrs
8Hours
20 Hours
150
200
250
350
400
455
Micro-seconds
0
100
200
300
400
500
600
700
800
900
1000
0 2 4 6 8 10
T
I
M
E
_
No_of_joins
EXA_ENU
GA_OA

6. CONSLUSION & FUTURE WORK
The aim of this research paper is limited to proposing a stochastic solution to the operation
allocation problem of Distributed Database Design. Most of the commercial vendors of Distributed
DBMS to date use exhaustive enumeration procedures along with different heuristics. They also
incorporate solutions based on other algorithm design techniques like Dynamic Programming,
Backtracking etc. This paper highlights that exhaustive procedures quickly go intractable when
No. of sites, or, No. of joins are increased suddenly. Exhaustive Enumerations along with
heuristics guarantee an optimal solution but total time of query is too large to be practically
viable.GA_OA does not guarantee the most optimal solution but provides very near to the best
solution in a very short span of time.
In future efforts should be done to incorporate Genetic Based Solutions to allocation problems of
Distributed Database. More work needed to be done to ensure that an optimal solution is
guaranteed in most of situations by GA’s. Furthermore Fragmentation, operation allocation and
Data Allocation and Load Balancing need to be integrated in one robust Genetic Solution.
2. REFERENCES
[1] Ozsu & Valduriez. “Principles of Distributed Database Systems” Pearson Education 2nd
Edition,pp. 228-298.
[2] March,Rho,”Characterisation and Analysis of a Nested Genetic Algorithm for Distributed
Database Design”.,Seoul Journal of Business pp 85-121 vol2,Number 1. 1995.
[3] Goldberg David.E “Genetic Algorithms in search, Optimization & Learning” Pearson
Education 2
nd
Edition,pp. 1-55.
[4] Sacco,G. & Yao”Query Optimisation in Distributed Database Systems”1982,Advances in
Computers,21,225-53.
[5] Yu,C.T,Chang” Distributed Query Processing “ ACM Computing Surveys,16,399-433.
[6] Graefe,G”Query Evalution Texhniques for a large Database” ACM Computing
Surveys,25,73-90, .1993.

[7] March,S.T.,Rho “Allocating Data and Operations to nodes in a distributed database
design”. IEEE Trans. On knowledge and Data Engg.,7(2). 1995.
[8] Kossman,D. “The state of the art in Distributed Query Processing”.,ACM Computing
Surveys.,32(4),422-469. 2000.
[9] Cheng,C.H.Lee,W-K,Wong,K-F, “A Genetic Agorithm based clustering approach for
database partitioning “ IEEE Transactions on System,Man,Cybernetics,32(3),215-230.
2002.
[10] Zehai Zhou,”Using Heuristics and Genetic Algorithms for Large Scale Database
Query Optimization,” Journal of Information and Computing Sciences,Acadeamim Press-
2007.
[11] Apers,P.M.G,1988”Data Allocation in Distributed Database Systems”,ACM Trans. On
Database Syatems,.13(3),263-304
[12] Tamhankar,A.M & Ram”Database Fragmentation & Allocation: An Integrated Methodolgy
and case study.” IEEE Transactions on System,Man,Cybernetics,28(3),288-305.
[13] Martin,T,Lam& Russel”An Evaluation of Site Selection Algorithms for Distributed Query
Processing”’The Compuer Journal,33(1),61-70,1990.
[14] Frieder, O. Baru”Site and Quey Sechduling policies in Microcomputer Database Systems”
IEEE Trans. On knowledge and Data Engg.,6(4).1994.
[15] Johansson,JM,March,ST,Naumann”Modelling Network Latency Paralell Processing In
Distributed Database design”,Decision Sciences,34(4) 677-706 2003.

Optimized Access Strategies for a Distributed Database Design

More Related Content

What's hot (20)

Viewers also liked (18)

Similar to Optimized Access Strategies for a Distributed Database Design (20)

More from Waqas Tariq (20)

Recently uploaded (20)

Optimized Access Strategies for a Distributed Database Design