Transcript Hipc-poster-final

Information Representation on the Grid: a Synthetic Database Benchmark/Workload for Grid
Information Servers
Beth Plale, Craig Jacobs, Ying Liu, Charlie Moad, Rupali Parab, Prajakta Vaidya, Nithya Vijayakumar
Indiana University, Bloomington, USA
Motivation
Query Results Analysis
The Grid is new paradigm for wide area distributed computing. Management of
grid resource information is complicated by fact that grid resources such as
hosts, clusters, virtual organizations, people, mathematical libraries, software
packages and services have aspects of their description that change at
millisecond rates. This defining characteristic makes traditional directory
service solutions inappropriate.
Contribution
•Understanding of resource information representation and retrieval in grid
computing through development of grid-specific synthetic database
benchmark/workload for grid resource information repository.
•Application of database benchmark on three vastly heterogeneous
“database” platforms: mySQL 4.0 (RDBMS), Xindice 1.1 (XML database) and
MDS2 (LDAP database)
•Scoped limits a search to a particular subtree. It appears to achieve its effect
when collection size is large.
•Selectivity is number of objects satisfying a query. From figures, we can see
MDS2 is sensitive to returned set size.
•Indexed queries over an Indexed attribute. Comparison of Indexed and nonIndexed queries for relational and XML platforms reveals obvious conclusion
that indexes greatly improve performance.
•Joins: “manyRelations” and “jobSubmit” join queries measure repository’s
ability to assemble result from numerous collections. In comparing join and
selectivity results for Xindice, join is much faster than selectivity. This is
because selectivity queries is over very large collection. So we can conclude
that Xindice is quite sensitive to collection size.
•Performance metric that captures both tangible and intangible aspects of
information retrieval: Response time for Queries and Scenarios, Ease of Use.
•Update is a simple, one-attribute update. mySQL can do 200 updates per
second whereas Xindice can perform only 1.3 updates per second.
Grid Resource Benchmark/Workload Description
Database Creation
and Population
HPC site
Jan ’02,
NASA
Nov ‘00
Perl scripts
LDIF
file
LDAP
(MDS)
relational
(mySQL)
mySQL
dump
XML
(Xindice)
Collection
Cluster
Scenario Results
Number of Objects
20
User Accounts
60
Computing Elements
110
Sub Clusters
340
Application
600
Connection
12000
Host
29520
Queries/updates are grouped into five categories:
• Scoped – restrict query evaluation to a particular sub-tree
• Indexed – request indexed attributes;
• Selectivity – number of objects returned;
• Joins – ask a realistic grid related question;
• Update/Connection – update single attribute; simple connection request
Scenarios are scripted synthetic workload issued over controlled time duration
consisting of concurrent query and update/insert requests. Scenarios consist
of three phases:
• Phase I (6 minutes) – Execution of several concurrent query streams.
• Phase II (5 minutes) – Update threads started along with query threads started
in Phase I.
• Phase III (9 minutes) – Update threads stopped. Query streams of Phase I still
running.
Ease of Use
Ease of Use captures intangible aspects of performance of a grid service, in
particular, amount of work client must undertake to obtain desired
information. That is, number of queries user must issue and amount of
processing required on returned data that falls upon user.
Metrics used to quantify ease of use are number of bytes returned and
number of queries needed in order to retrieve requested data. Database
platforms requiring more queries to obtain data correspondingly return larger
number of bytes.
Description
mySQL 4.0
(KB)
Xindice 1.1
(KB)
MDS2
(KB)
Description
mySQL 4.0
Xindice 1.1
MDS2
Scoped
0.4 – 46.0
7.5 – 549.5
5.6 – 47.3
Scoped
1
3
3
Indexed
9.5 – 11.0
139.8 – 140.9
9.6 – 24.0
Indexed
1
2
1
Selectivity
0.04 – 52.9
0.48 – 691.0
0.03 – 267.0
Selectivity
1
1
1
Joins
0.03 – 0.03
40.1 – 131.8
0.98 – 1.9
Joins
1
6
5
Release of IURGRbench benchmark v1.0 is currently available at
website: http://www.cs.indiana.edu/~plale/RGR/