You are here

Ease-of-Use Focused Performance Modeling of Distributed Enterprise Systems

Speaker(s): 
Presentation Type: 
talk

In this talk I will be describing my work on performance modeling, which I have been conducting in partnership with IBM Hursley and IBM Research. This focuses on making performance modeling easier to use by users with limited modeling expertise, in the domain of distributed enterprise systems. In particular, I am interested in the following performance modeling requirements: 1.) giving the modeler a choice of performance modeling methods; 2.) allowing the modeler to model a wide range of systems within the domain; and make predictions using a range of performance metrics; 3.) having models that can be evaluated rapidly but accurately (e.g. in seconds on one core/processor); and 4.) having models that can be parametrised quickly and at a low monitoring overhead (e.g. on a live system).

So far I have been focusing on using and comparatively evaluating three performance modeling methods: the Layered Queuing Method; statistically extrapolating from historical performance data (using my own method); and a hybrid of the two - although I have also been considering a wider range of methods. The talk will also describe some of the extensive experimental results I have obtained to validate my work. These are supported by measured performance data from an in-house copy of the hardware/software testbed IBM Hursley use to test much of their software. Finally, I will touch upon some of the applications of this work (e.g. Grid scheduling).

Date and time: 
Tuesday, 20 January, 2009 - 16:00
Location: 
IF3.02, Informatics PEPA-club