[Colloquium] Raicu/M.S. Presentation/May 6, 2005

[Margaret Jaffey]

This is an announcement of Ioan Raicu's Master's Presentation.

------------------------------------------------------------------------ 
---
Date:  Friday, May 6, 2005

Time:  1:30 p.m.

Place:  Ryerson 276

M.S. Candidate:  Ioan Raicu

M.S. Paper Title:  A Performance Study of the Globus Toolkit® and Grid  
Services via DiPerF, an automated DIstributed PERformance testing  
Framework

Abstract:
The Globus Toolkit® (GT®) is the “de facto standard” for grid  
computing. Measuring the performance of various components of the GT®  
in a wide area network (WAN) as well as a local area network (LAN) is  
essential in understanding the performance that is to be expected from  
the GT® in a realistic deployment in a distributed and heterogeneous  
environment. Furthermore, measuring the performance of grid services in  
a WAN is similarly important due to the complex interactions between  
network connectivity and service performance.
Performing distributed measurements is not a trivial task, due to  
difficulties 1) accuracy – synchronizing the time across an entire  
system that might have large communication latencies, 2) flexibility –  
in heterogeneity normally found in WAN environments and the need to  
access large number of resources, 3) scalability – the coordination of  
large amounts of resources, and 4) performance – the need to process  
large number of transactions per second. In attempting to address these  
four issues, I developed DiPerF, a DIstributed PERformance testing  
Framework, aimed at simplifying and automating service performance  
evaluation. DiPerF coordinates a pool of machines that test a single or  
distributed target service, collects and aggregates performance metrics  
from the client point of view, and generates performance statistics.  
The aggregate data collected provides information on service  
throughput, service response time, on service ‘fairness’ when serving  
multiple clients concurrently, and on the impact of network latency on  
service performance.
DiPerF is a modularized tool, with various components written in  
C/C++/perl; it has been tested over PlanetLab, Grid3, and the Computer  
Science Cluster at the University of Chicago. I have implemented  
several variations (ssh, TCP, UDP) on the communication between  
components in order to give DiPerF the most flexibility and scalability  
in a wide range of scenarios. I have investigated both the performance  
and scalability of DiPerF and found that DiPerF is able to handle up to  
10,000+ clients and 100,000+ transactions per second; I also performed  
a validation study of DiPerF to ensure that the aggregate client view  
matched the tested service view and found that the two views usually  
matched within a few percent.
In profiling the performance of the GT®, we investigated the  
performance of three main components: 1) job submission, 2) information  
services, and 3) a file transfer protocol. All these components of the  
Globus Toolkit® are vital to the functionality, performance, and  
scalability of the grid. We specifically tested: 1) pre WS GRAM and WS  
GRAM included with GT® 3.2 and 3.9.4, 2) the scalability and  
performance of the WS-MDS Index bundled with GT® 3.9.5, and 3) the  
scalability and fairness of the GridFTP server included with the GT®  
3.9.5. We also investigated the performance of two grid services: 1)  
DI-GRUBER, a distributed usage SLA-based broker, a grid service based  
on the Globus Toolkit® 3.2 and 3.9.5, and 2) instance creation and  
message passing performance in the Globus Toolkit® 3.2.
Through the various test case studies I performed using DiPerF, I  
believe it has proved to be a valuable tool for scalability and  
performance evaluation studies, as well as for automated extraction of  
service performance characteristics. We conclude with future work that  
could benefit from the foundation established by DiPerF; we outline how  
we believe the performance models that can be extracted from DiPerF can  
be used to perform predictive scheduling in the Grid.

Advisor:  Prof. Ian Foster

A draft copy of Ioan Raicu's MS Paper will be available soon in Ry 161A.

--  
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-
Margaret P. Jaffey				margaret at cs.uchicago.edu
Department of Computer Science
Student Support Rep (Ry 161A)		(773) 702-6011
The University of Chicago		http://www.cs.uchicago.edu
=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-