SIGMOD 99 Efficient Concurrency Control in Multidimensional Access Methods Kaushik Chakrabarti Sharad Mehrotra University of.

SIGMOD 99 http://www-db.ics.uci.edu

Efficient Concurrency Control in Multidimensional Access Methods

Kaushik Chakrabarti Sharad Mehrotra

University of Illinois at Urbana Champaign

University of California at Irvine

Presented at ACM SIGMOD Conference

June 1, 1999

SIGMOD 99 http://www-db.ics.uci.edu

Outline of talk

• Introduction

• Background

• Phantom protection in Generalized Search Trees– Define granules– Describe lock protocols

• Experiments

• Conclusion

SIGMOD 99 http://www-db.ics.uci.edu

Introduction• Increasing number of applications deal with multidimensional data

• Examples: spatial (CAD, GIS), spatio-temporal (moving objects, weather)

• DBMS should allow applications to:(1) define their own data types and operations(2) define multidimensional access methods (AMs) for those data types for efficient

query processing

• OR technology solves (1)

• Generalized Search Trees (GiSTs) addresses (2)

SIGMOD 99 http://www-db.ics.uci.edu

Introduction • For successful integration, we need to support concurrent accesses via

GiST

• Concurrency control problems:(1) Preserve consistency of data structure(2) Prevent phantom anomalies

• (1) has been addressed in Kornacker, Mohan and Hellerstein, SIGMOD97

• This paper addresses the problem of phantom protection in GiSTs

SIGMOD 99 http://www-db.ics.uci.edu

Phantom

• Definition:– T1 reads a set of items satisfying some <search-condition>

– T2 creates data items that satisfy T1’s <search-condition> and commits

– T1 repeats its scan with the same <search-condition>,

gets a different set of items

• Serializability No phantoms

SIGMOD 99 http://www-db.ics.uci.edu

Example

SIGMOD 99 http://www-db.ics.uci.edu

Solution

• Predicate locks: costly

• Granular locks: efficient

SIGMOD 99 http://www-db.ics.uci.edu

Key Range Locking

• ARIES/KVL(Mohan, 1990)

SIGMOD 99 http://www-db.ics.uci.edu

Phantoms in Spatial/Spatio-temporal Databases

• Compute average rainfall over all locations a 2-d region where the locations are indexed using a GiST

• Get all objects in a given region from a moving objects database where the objects are indexed using a GiST

SIGMOD 99 http://www-db.ics.uci.edu

Solutions

• Adapting KRL: too costly.

• Predicate locking based strategy by Kornacker, Mohan and Hellerstein, SIGMOD97.

• Our granular locking based approach for phantom protection in R-trees, ICDE98. Does not work well when applied to GiSTs (details in paper)

SIGMOD 99 http://www-db.ics.uci.edu

Granular Locking in GiST

• Solution involves– Define the granules– Define the lock protocol for the operations

• Challenges– “nice’’ granules– handling overlap among granules– handling “loss of lock’’ problem– high concurrency and low lock overhead

SIGMOD 99 http://www-db.ics.uci.edu

GiST

• Keys can be arbitrary predicates• An AM can be implemented by specifying some

extension methods which dictate the tree operations

SIGMOD 99 http://www-db.ics.uci.edu

Granules in GiST

• Leaf Granules: One per leaf node

• Non-leaf granules: One per non-leaf node

• Lock name: <table-name, index-name, node-id>

• Lock Coverage: defined by Granule Predicate (GP)

– GP(N) = BP(N) if N is root

= BP(N) GP(P) otherwise, P=parent(N)

SIGMOD 99 http://www-db.ics.uci.edu

SIGMOD 99 http://www-db.ics.uci.edu

Locks

SIGMOD 99 http://www-db.ics.uci.edu

Overlap between granules• Correctness: p p’ lset(p) lset(p’) NULL

• Problem does not arise in KRL

• Policies– Overlap-for-Search & Cover-for-Insert (OSCI)– Cover-for-Search & Overlap-for-Insert (CSOI)

SIGMOD 99 http://www-db.ics.uci.edu

Loss of lock coverage

SIGMOD 99 http://www-db.ics.uci.edu

Search Protocol

• Get commit duration S lock on the granule corresponding to each index node visited

• Correctness:– GP(T) Q is satisfiable i(Consistent(BP(Pi), Q), Pi is

ancestor of T

• Note– No object locks– No extra cost except that of acquiring the lock (no extra checks)

SIGMOD 99 http://www-db.ics.uci.edu

Insert Protocol

• Correctness:– full coverage– prevent phantoms due to loss of lock

coverage

SIGMOD 99 http://www-db.ics.uci.edu

Insert Protocol

• Case 1: No growth, No split– commit duration IX lock on g (target granule)– commit duration X lock on O

• Case 2: Growth, No split– 2 locks as before– short duration IX lock on lowest unchanged node

(LU-node)

SIGMOD 99 http://www-db.ics.uci.edu

Example

SIGMOD 99 http://www-db.ics.uci.edu

Insert Protocol

• Case 3: No growth, Split– instant duration SIX on g– commit duration IX on whichever contains O after

split; X on O– instant duration SIX on each ancestor that splits

• Case 4: Growth, Split– lock requirements of Cases 2 and 3

SIGMOD 99 http://www-db.ics.uci.edu

Deletion Protocol

• Problem: g does not cover O after deletion commit duration lock on LU-node

• We do: – logical deletion (IX on target granule, X on object)– defer physical deletion till transaction commits

SIGMOD 99 http://www-db.ics.uci.edu

Protocol for Other Operations

• ReadSingle: S lock on object

• UpdateSingle: – if indexed attributes not changed, IX on g, X on O– else, deletion followed by insertion

• UpdateScan: same as search for the region, same as updatesingle for every object updated

SIGMOD 99 http://www-db.ics.uci.edu

Empirical Evaluation

• Data sets:– 2-d spatial data: 62,556 2-d points from Sequoia

2000 benchmark

– 3-d feature data: First 3 Fourier coefficients from 480,471 Fourier vectors

SIGMOD 99 http://www-db.ics.uci.edu

Measurements & Parameters

• Performance: Throughput (tps)

• Concurrency: Conflict ratio

• Overhead: #locks, # pred. Checks

• Parameters: MPL, transaction size, write probability, query size, external think time (fixed 3sec), restart delay (fixed 3sec)

SIGMOD 99 http://www-db.ics.uci.edu

Implementation

SIGMOD 99 http://www-db.ics.uci.edu

Performance

• 2-d data

Throughput vs MPL

0

2

4

6

8

10

12

14

16

10 20 40 50 60 80 100

MPL

Thro

ughp

ut (t

ps) GL

PL

• 3-d data

Throughput vs MPL

0

1

2

3

4

5

6

7

8

9

10

10 30 40 50 60 75

MPL

Thro

ughp

ut (t

ps) GL

PL

SIGMOD 99 http://www-db.ics.uci.edu

Performance/Concurrency

• Under various loads

Throughput vs Write ratio

0

5

10

15

0.1 0.2 0.4 0.5 0.6 0.8 0.9

Write ratio

Throu

ghpu

t (tps

) GL

PL

• Conflict ratio

Concurrency vs. MPL

00.5

11.5

MPL

Confl

ict ra

tio GLPL

SIGMOD 99 http://www-db.ics.uci.edu

Overhead

• Search • Insert

Lock Overhead

0

50

100

150

200

250

MPL

# lo

cks/

# p

red

. ch

ecks GL

PL

Lock Overhead

0100200300400500600700800

MPL

# lo

cks/

# p

red

. ch

ecks GL

PL

SIGMOD 99 http://www-db.ics.uci.edu

Conclusions• GL is significantly more efficient than PL

• We expect the performance gap to increase with better implementation (mainly LM)

• Dimensionality curse is a problem in GL

• Can be integrated with a consistency protocol for complete solution to concurrency control in multidimensional AMs