Parallel Database

1.Parallel and Distributed Databases

1 1. Parallel DB /D.S.Jagli 2/12/2013
1. Parallel DB /D.S.Jagli 1

Parallel database
1. -Introduction
2. -Architecture for Parallel databases.
3. - Parallel query Evaluation
4. - Parallelizing Individual operations.


Introduction

 What is a Centralized Database ?
-all the data is maintained at a single site and assumed that the processing of
individual transaction is essentially sequential.


PARALLEL DBMSs
WHY DO WE NEED THEM?

• More and More Data!

We have databases that hold a high amount of
data, in the order of 1012 bytes:

10,000,000,000,000 bytes!

• Faster and Faster Access!

We have data applications that need to process
data at very high speeds:

10,000s transactions per second!

SINGLE-PROCESSOR DBMS AREN’T UP TO THE JOB!

Why Parallel Access To Data?
At 10 MB/s 1,000 x parallel
1.2 days to scan 1.5 minute to scan.

1 Terabyte 1 Terabyte

10 MB/s
Parallelism:
divide a big problem
1. Parallel DB /D.S.Jagli
into many smaller ones
5 2/12/2013
to be solved in parallel.

Parallel DB
 Parallel database system seeks to improve performance through
parallelization of various operations such as loading data ,building
indexes, and evaluating queries by using multiple CPUs and Disks in
Parallel.

 Motivation for Parallel DB
 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


PARALLEL DBMSs
BENEFITS OF A PARALLEL DBMS

 Improves Response Time.

INTERQUERY PARALLELISM
It is possible to process a number of transactions in
parallel with each other.

 Improves Throughput.

INTRAQUERY PARALLELISM
It is possible to process ‘sub-tasks’ of a transaction in
parallel with each other.


PARALLEL DBMSs
HOW TO MEASURE THE BENEFITS

 Speed-Up
– Adding more resources results in proportionally less running time for a
fixed amount of data.
10 seconds to scan a DB of 10,000 records using 1 CPU
1 second to scan a DB of 10,000 records using 10 CPUs

 Scale-Up
 If resources are increased in proportion to an increase in data/problem
size, the overall time should remain constant

– 1 second to scan a DB of 1,000 records using 1 CPU
1 second to scan a DB of 10,000 records using 10 CPUs


Architectures for Parallel Databases
 The basic idea behind Parallel DB is to carry out evaluation steps in
parallel whenever is possible.

 There are many opportunities for parallelism in RDBMS.

 3 main architectures have been proposed for building parallel DBMSs.
1. Shared Memory
2. Shared Disk
3. Shared Nothing


Shared Memory
 Advantages:
1. It is closer to conventional
machine , Easy to program
2. overhead is low.
3. OS services are leveraged to
utilize the additional CPUs.
 Disadvantage:
1. It leads to bottleneck problem
2. Expensive to build
3. It is less sensitive to
partitioning


Shared Disk
 Advantages:
1. Almost same
 Disadvantages:
1. More interference
2. Increases N/W band width
3. Shared disk less sensitive to
partitioning


Shared Nothing
 Advantages:
1. It provides linear scale up
&linear speed up
2. Shared nothing benefits from
"good" partitioning
3. Cheap to build
 Disadvantage
1. Hard to program
2. Addition of new nodes
requires reorganizing


PARALLEL DBMSs
SPEED-UP

Number of transactions/second

Linear speed-up (ideal)

2000/Sec
1600/Sec
Sub-linear speed-up
1000/Sec

5 CPUs 10 CPUs 16 CPUs

Number of CPUs 2/12/2013
13

PARALLEL DBMSs
SCALE-UP

Number of transactions/second

1000/Sec Linear scale-up (ideal)
900/Sec Sub-linear scale-up

5 CPUs 10 CPUs
1 GB Database 2 GB Database

Number of CPUs, Database size 2/12/2013
14

PARALLEL QUERY EVALUATION
A relational query execution plan is graph/tree of
relational algebra operators (based on this operators can
execute in parallel)

1. 1. Parallel DB /D.S.Jagli 15

Different Types of DBMS ||-ism
 Parallel evaluation of a relational query in DBMS With shared –nothing
architecture
1. Inter-query parallelism
 Multiple queries run on different sites
2. Intra-query parallelism
 Parallel execution of single query run on different sites.
a) Intra-operator parallelism
a) get all machines working together to compute a given operation (scan, sort,
join).
b) Inter-operator parallelism
 each operator may run concurrently on a different site (exploits
pipelining).
 In order to evaluate different operators in parallel, we need to
evaluate each operator in query plan in Parallel.
1. 1. Parallel DB /D.S.Jagli 16

Data Partitioning
 Types of Partitioning
1. Horizontal Partitioning: tuple of a relation are divided among
many disks such that each tuple resides on one disk.
 It enables to exploit the I/O band width of disks by reading & writing
them in parallel.
 Reduce the time required to retrieve relations from disk by
partitioning the relations on multiple disks.
1. Range Partitioning
2. Hash Partitioning
3. Round Robin Partitioning

2. Vertical Partitioning
2/12/2013
17 1. Parallel DB /D.S.Jagli
17

1.Range Partitioning
 Tuples are sorted (conceptually), and n ranges are chosen for
the sort key values so that each range contains roughly the
same number of tuples;
 tuples in range i are assigned to processor i.
 Eg:
 sailor _id 1-10 assigned to disk 1

 range partitioning can lead to data skew; that is, partitions with widely
varying number of tuples across

18

2.Hash Partitioning
 A hash function is applied to selected fields of a tuple to determine its
processor.

 Hash partitioning has the additional virtue that it keeps data evenly
distributed even if the data grows and shrinks over time.

19

3.Round Robin Partitioning

 If there are n processors, the i th tuple is assigned to processor i mod n in
round-robin partitioning.

 Round-robin partitioning is suitable for efficiently evaluating queries that
access the entire relation.

 If only a subset of the tuples (e.g., those that satisfy the selection
condition age = 20) is required, hash partitioning and range partitioning
are better than round-robin partitioning

20

Range Hash Round Robin

A...E F...J K...N O...S T...Z A...E F...J K...N O...S T...Z A...E F...J K...N O...S T...Z

Good for equijoins, Good for equijoins, Good to spread load
exact-match queries, exact match queries
and range queries


Parallelizing Sequential Operator
Evaluation Code
1. An elegant software architecture for parallel DBMSs enables us to
readily parallelize existing code for sequentially evaluating a
relational operator.

2. The basic idea is to use parallel data streams.

3. Streams are merged as needed to provide the inputs for a relational
operator.

4. The output of an operator is split as needed to parallelize subsequent
processing.

5. A parallel evaluation plan consists of a dataflow network of
relational, merge, and split operators.

PARALLELIZING INDIVIDUAL
OPERATIONS
 How various operations can be implemented in parallel in a shared-
nothing architecture?

 Techniques
1. Bulk loading& scanning
2. Sorting
3. Joins


1.Bulk Loading and scanning
 scanning a relation: Pages can be read in parallel while scanning a
relation, and the retrieved tuples can then be merged, if the relation is
partitioned across several disks.

 bulk loading: if a relation has associated indexes, any sorting of data
entries required for building the indexes during bulk loading can also
be done in parallel.


2.Parallel Sorting :
 Parallel sorting steps:
1. First redistribute all tuples in the relation using range partitioning.
2. Each processor then sorts the tuples assigned to it
3. The entire sorted relation can be retrieved by visiting the processors in
an order corresponding to the ranges assigned to them.

 Problem: Data skew

 Solution: “sample” the data at the outset to determine good
range partition points.

A particularly important application of parallel sorting is sorting the data
entries in tree-structured indexes.

3.Parallel Join
1. The basic idea for joining A and B in parallel is to decompose the join
into a collection of k smaller joins by using partition.

2. By using the same partitioning function for both A and B, we ensure that
the union of the k smaller joins computes the join of A and B.

 Hash-Join
 Sort-merge-join


Sort-merge-join
 partition A and B by dividing the range of the join attribute into k disjoint
subranges and placing A and B tuples into partitions according to the
subrange to which their values belong.

 Each processor carry out a local join.

 In this case the number of partitions k is chosen to be equal to the number
of processors n .

 The result of the join of A and B, the output of the join process may be
split into several data streams.

The advantage that the output is available in sorted order

Dataflow Network of Operators for
Parallel Join

Good use of split/merge makes it easier to build parallel versions of sequential
join code

DDBMS

I. Introduction to DDBMS
II. Architecture of DDBs
III. Storing data in DDBs
IV. Distributed catalog management
V. Distributed query processing
VI. Transaction Processing
VII. Distributed concurrency control and recovery

I.Introduction to DDBMS
 Data in a distributed database system is stored across several sites.

 Each site is typically managed by a DBMS that can run independently of
the other sites that co-operate in a transparent manner.

 Transparent implies that each user within the system may access all of
the data within all of the databases as if they were a single database

 There should be „location independence‟ i.e.- as the user is unaware of
where the data is located it is possible to move the data from one physical
location to another without affecting the user.


DDBMS properties
 Distributed data independence: Users should be able to ask queries
without specifying where the referenced relations, or copies or
fragments of the relations, are located.

 Distributed transaction atomicity: Users should be able to write
transactions that access and update data at several sites just as they
would write transactions over purely local data.


DISTRIBUTED PROCESSING ARCHITECTURE

CLIENT CLIENT
CLIENT CLIENT

LAN
LAN
CLIENT CLIENT
CLIENT CLIENT

Delhi Mumbai

CLIENT CLIENT CLIENT CLIENT

DBMS
LAN LAN
CLIENT CLIENT
CLIENT CLIENT

32 1. Parallel DB /D.S.Jagli
Hyderabad Pune2/12/2013

DISTRIBUTED DATABASE ARCHITECTURE


DBMS
DBMS

LAN


Delhi Mumbai

CLIENT

DBMS
DBMS

LAN


Hyderabad DB /D.S.Jagli
33 1. Parallel
Pune
2/12/2013
33

Distributed database
 Communication Network- DBMS and Data at each node

•Users are unaware
of the distribution of
the data
Location
transparency


Types of Distributed Databases
 Homogeneous distributed database system :
 If data is distributed but all servers run the same DBMS software.

 Heterogeneous distributed database :
 If different sites run under the control of different DBMSs,
essentially autonomously, are connected to enable access to data
from multiple sites.

 The key to building heterogeneous systems is to have well-accepted
standards for gateway protocols.
 A gateway protocol is an API that exposes DBMS functionality to
external applications.


DDBMS


2.DISTRIBUTED DBMS
ARCHITECTURES
1. Client-Server Systems:
2. Collaborating Server Systems
3. Middleware Systems


1.Client-Server Systems:
1. A Client-Server system has one or more client
processes and one or more server processes,
2. A client process can send a query to any one server
process.
3. Clients are responsible for user-interface issues,
4. Servers manage data and execute transactions.
5. A client process could run on a personal computer
and send queries to a server running on a mainframe.
6. The Client-Server architecture does not allow a single
query to span multiple servers


SPECIALISED NETWORK CONNECTION
TERMINALS
MAINFRAME COMPUTER
DUMB

DUMB

DUMB PRESENTATION LOGIC
BUSINESS LOGIC
DATA LOGIC

2.Collaborating Server Systems
1. The client process would have to be capable of breaking
such a query into appropriate subqueries.
2. A Collaborating Server system can have a collection of
database servers, each capable of running transactions
against local data, which cooperatively execute
transactions spanning multiple servers.
3. When a server receives a query that requires access to
data at other servers, it generates appropriate subqueries
to be executed by other servers .
4. puts the results together to compute answers to the
original query.


M:N CLIENT/SERVER DBMS ARCHITECTURE

SERVER #1
CLIENT
#1
D/BASE

CLIENT
#2

SERVER #2

D/BASE
CLIENT
#3

NOT TRANSPARENT!

3.Middleware Systems:
 The Middleware architecture is designed to allow a single
query to span multiple servers, without requiring all
database servers to be capable of managing such multisite
execution strategies.

 It is especially attractive when trying to integrate several
legacy systems, whose basic capabilities cannot be
extended.


DDBMS


3.Storing Data in DDBs
 In a distributed DBMS, relations are stored across
several sites.
 Accessing a relation that is stored at a remote site
includes message-passing costs.
 A single relation may be partitioned or fragmented
across several sites.


Types of Fragmentation:
Horizontal fragmentation: The union of the horizontal
fragments must be equal to the original relation. Fragments
are usually also required to be disjoint.

Vertical fragmentation: The collection of vertical fragments
should be a lossless-join decomposition.

45

46

Replication
 Replication means that we store several copies of a relation
or relation fragment.

 The motivation for replication is twofold:
1. Increased availability of data:
2. Faster query evaluation:

 Two kinds of replications
1. synchronous replication
2. asynchronous replication

47

DDBMS

IV. Distributed Catalog Management
V. Distributed Query Processing

4.Distributed Catalog Management
1. Naming Objects
• If a relation is fragmented and replicated, we must be able to
uniquely identify each replica of each fragment.
1. A local name field
2. A birth site field

2. Catalog Structure
 A centralized system catalog can be used It is vulnerable to failure of
the site containing the catalog).
 An alternative is to maintain a copy of a global system catalog.
compromises site autonomy,)

49

4.Distributed Catalog Management
 A better approach:
 Each site maintains a local catalog that describes all copies
of data stored at that site.

 In addition, the catalog at the birth site for a relation is
responsible for keeping track of where replicas of the
relation are stored.

50

DDBMS

IV. Distributed Catalog management
V. Distributed Query Processing

5.Distributed Query Processing
 Distributed query processing: Transform a high-level
query (of relational calculus/SQL) on a distributed database
(i.e., a set of global relations) into an equivalent and
efficient lower-level query (of relational algebra) on
relation fragments.

 Distributed query processing is more complex
1. – Fragmentation/replication of relations
2. – Additional communication costs
3. – Parallel execution

52

Distributed Query Processing Steps


 Sailors(sid: integer, sname: string, rating: integer, age: real)
 Reserves(sid: integer, bid: integer, day: date, rname: string)

 Assume Reserves and Sailors relations
 each tuple of Reserves is 40 bytes long
 a page can hold 100 Reserves tuples
 1,000 pages of such tuples.
 each tuple of Sailors is 50 bytes long
 a page can hold 80 Sailors Tuples
 500 pages of such tuples

 How to estimate the cost?

54

 Criteria for measuring the cost of a query evaluation
strategy
 For centralized DBMSs number of disk accesses (#
blocks read / written)
 For distributed databases, additionally
 The cost of data transmission over the network
 Potential gain in performance from having several sites processing parts
of the query in parallel


5.Distributed query processing
 To estimate the cost of an evaluation strategy, in addition to counting
the number of page I/Os.
 we must count the number of pages that are shipped is a
communication costs.
 Communication costs is a significant component of overall cost in a
distributed database.

1. Nonjoin Queries in a Distributed DBMS
2. Joins in a Distributed DBMS
3. Cost-Based Query Optimization

56

1.Nonjoin Queries in a Distributed DBMS
 Even simple operations such as scanning a relation, selection, and
projection are affected by fragmentation and replication.

SELECT S.age
FROM Sailors S
WHERE S.rating > 3 AND S.rating < 7
 Suppose that the Sailors relation is horizontally fragmented, with all
tuples having a rating less than 5 at Mumbai and all tuples having a
rating greater than 5 at Delhi.

The DBMS must answer this query by evaluating it at both sites and
taking the union of the answers.

2/12/2013 57

Eg 1: SELECT avg(age)
FROM Sailors S
WHERE S.rating > 3 AND S.rating < 7
 taking the union of the answers is not enough
Eg 2: SELECT S.age
FROM Sailors S
WHERE S.rating > 6
 taking the union of the answers is not enough

 Eg 3: suppose that the Sailors relation is vertically fragmented, with the
sid and rating fields at MUMBAI and the sname and age fields at
DELHI
 This vertical fragmentation would be a lossy decomposition


Eg 4:the entire Sailors relation is stored at both MUMBAI and DELHI
sites.
 Where should the query be executed?
 2 Joins in a Distributed DBMS
 Eg: the Sailors relation is stored at MUMBAI, and that the Reserves
relation is stored at DELHI.
 Joins of relations at different sites can be very expensive.
 JOIN STRATEGY
1. Fetch as needed
2. Ship whole
3. Semijoins
4. Bloomjoins

59 1. Parallel DB /D.S.Jagli Which strategy is better for me? 2/12/2013

 1.Fetch As Needed
 Page-oriented Nested Loops join: For each page of R, get each page
of S, and write out matching pairs of tuples <r, s>, where r is in R-page
and s is in S-page.

 We could do a page-oriented nested loops join in MUMBAI with
Sailors as the outer, and for each Sailors page, fetch all Reserves pages
from DELHI.

 If we cache the fetched Reserves pages in MUMBAI until the join is
complete, pages are fetched only once


Fetch As Needed: Transferring the relation
piecewise

QUERY: The query asks for R S


 Assume Reserves and Sailors relations
 each tuple of Reserves is 40 bytes long
 a page can hold 100 Reserves tuples
 1,000 pages of such tuples.
 each tuple of Sailors is 50 bytes long
 a page can hold 80 Sailors Tuples
 500 pages of such tuples
 td is cost to read/write page; ts is cost to ship page.
 The cost is = 500td to scan Sailors
 for each Sailors page, the cost of scanning shipping all of Reserves, which is
=1000(td + ts).
 The total cost is = 500td + 500, 000(td + ts).


 This cost depends on the size of the result.
 The cost of shipping the result is greater than the cost of shipping both
Sailors and Reserves to the query site.

 2.Ship to One Site (sort-merge-join)
 The cost of scanning and shipping Sailors, saving it at DELHI, and then
doing the join at DELHI is

=500(2td + ts) + (Result)td,

 The cost of shipping Reserves and doing the join at MUMBAI is

=1000(2td+ts)+(result)td.


Ship Whole: Transferring the complete
relation

QUERY: The query asks for R S


 Ship Whole vs Fetch as needed:
 Fetch as needed results in a high number of messages
 Ship whole results in high amounts of transferred data
 Note: Some tuples in Reserves do not join with any tuple in Sailors, we
could avoid shipping them.
3.Semijoins and Bloomjoins
Semijoins: 3 steps:
1. At MUMBAI, compute the projection of Sailors onto the join columns i.e
sid and ship this projection to DELHI.
2. 2. At DELHI, compute the natural join of the projection received from
the first site with the Reserves relation.
3. The result of this join is called the reduction of Reserves with respect to
Sailors. ship the reduction of Reserves to MUMBAI.
4. 3. At MUMBAI, compute the join of the reduction of Reserves with
65 Sailors. DB /D.S.Jagli
1. Parallel 2/12/2013

Semijoin:
 Semijoin: Requesting all join partners in just one step.


5.Distributed query processing
 Bloomjoins: 3 steps:
1. At MUMBAI, A bit-vector of (some chosen) size k is computed by
hashing each tuple of Sailors into the range 0 to k −1 and setting bit i to 1.
if some tuple hashes to i, and 0 otherwise then ship this to DELHI

2. At DELHI, the reduction of Reserves is computed by hashing each tuple
of Reserves (using the sid field) into the range 0 to k − 1, using the same
hash function used to construct the bit-vector, and discarding tuples
whose hash value i corresponds to a 0 bit.ship the reduction of Reserves
to MUMBAI.

3. At MUMBAI, compute the join of the reduction of Reserves with Sailors.


Bloom join:
 Bloom join:
 Also known as bit-vector join
 Avoiding to transfer all join attribute values to the other
node
 Instead transferring a bitvector B[1 : : : n]
 Transformation
 Choose an appropriate hash function h
 Apply h to transform attribute values to the range [1 : : : n]
 Set the corresponding bits in the bitvector B[1 : : : n] to


Bloom join:


 Cost-Based Query Optimization
 optimizing queries in a distributed database poses the following additional
challenges:

 Communication costs must be considered. If we have several copies of a
relation, we must also decide which copy to use.

 If individual sites are run under the control of different DBMSs, the
autonomy of each site must be respected while doing global query
planning.


Cost-Based Query Optimization
 Cost-based approach; consider all plans, pick cheapest; similar to
centralized optimization.

 Difference 1: Communication costs must be considered.
 Difference 2: Local site autonomy must be respected.
 Difference 3: New distributed join methods.

 Query site constructs global plan, with suggested local plans describing
processing at each site.
 If a site can improve suggested local plan, free to do so.


6.DISTRIBUTED TRANSACTIONS PROCESSING
 A given transaction is submitted at some one site, but it can
access data at other sites.

 When a transaction is submitted at some site, the transaction
manager at that site breaks it up into a collection of one or
more subtransactions that execute at different sites.




7.Distributed Concurrency Control
 Lock management can be distributed across sites in many
ways:
1. Centralized: A single site is incharge of handling lock and
unlock requests for all objects.
2. Primary copy: One copy of each object is designated as the
primary copy.
 All requests to lock or unlock a copy of this object are handled
by the lock manager at the site where the primary copy is stored.

3. Fully distributed: Requests to lock or unlock a copy of an
object stored at a site are handled by the lock manager at the
site where the copy is stored.

7.Distributed Concurrency Control
 Distributed Deadlock Detection:
 Each site maintains a local waits-for graph.
 A global deadlock might exist even if the local graphs contain no cycles:

 Three solutions:
 Centralized: send all local graphs to one site ;
 Hierarchical: organize sites into a hierarchy and send local graphs to parent
in the hierarchy;
 Timeout : abort Xact if it waits too long.


 Phantom Deadlocks: delays in propagating local information might
cause the deadlock detection algorithm to identify `deadlocks' that do not
really exist. Such situations are called phantom deadlocks


7.Distributed Recovery
 Recovery in a distributed DBMS is more complicated than in a
centralized DBMS for the following reasons:

1. New kinds of failure can arise:
 failure of communication links.
 failure of a remote site at which a subtransaction is
executing.
2. Either all subtransactions of a given transaction must
commit, or none must commit.
3. This property must be guaranteed in spite of any
combination of site and link failures.


7.Distributed Recovery
 Two-Phase Commit (2PC):

 Site at which Xact originates is coordinator;
 other sites at which it executes subXact are subordinates.
 When the user decides to commit a transaction:
 The commit command is sent to the coordinator for the
transaction.
 This initiates the 2PC protocol:


7.Distributed Recovery(2pc)
1. Coordinator sends prepare msg to each subordinate.

2. Subordinate force-writes an abort or prepare log record and then
sends a no or yes msg to coordinator.

3. If coordinator gets all yes votes, force-writes a commit log record and
sends commit msg to all subs. Else, force-writes abort log rec, and
sends abort msg.

4. Subordinates force-write abort/commit log rec based on msg they
get, then send ack msg to coordinator.

5. Coordinator writes end log rec after getting ack msg from all subs

TWO-PHASE COMMIT (2PC) - commit


TWO-PHASE COMMIT (2PC) - ABORT


7.Distributed Recovery(3pc)
 Three-Phase Commit
1. A commit protocol called Three-Phase Commit (3PC) can avoid
blocking even if the coordinator site fails during recovery.

2. The basic idea is that when the coordinator sends out prepare
messages and receives yes votes from all subordinates.

3. It sends all sites a precommit message, rather than a commit message.
4. When a sufficient number more than the maximum number of failures
that must be handled of acks have been received.
5. The coordinator force-writes a commit log record and sends a commit
message to all subordinates.


Distributed Database
 Advantages:  Disadvantages:
 Reliability  Complexity of Query
 Performance opt.
 Growth (incremental)  Concurrency control
 Local control  Recovery
 Transparency  Catalog management


Thank you

Queries?

Parallel Database

More Related Content

What's hot (20)

Similar to Parallel Database (20)

More from VESIT/University of Mumbai (10)

Recently uploaded (20)

Parallel Database