AESOP home

Publications

Scalable Performance Analysis of Massively Parallel Stochastic Systems

Richard Hayden

Seminar
Seminar Series on Advances in Telecommunications, Networking and Computing, Budapest University of Technology and Economics (BME)
November, 2012
Abstract

Performance analysis has always suffered from the state-space explosion problem which directly prohibits the scalability of stochastic modelling as a tool for resolving resource provisioning and quality of service questions in massively parallel computer and communication systems. This is especially true when applied to the recent ubiquitous breed of distributed and peer-to-peer systems.

One way around these scalability limitations are asymptotic techniques formally justified by functional laws of large numbers often termed variously ``fluid'' or ``mean-field'' analysis. These techniques have their roots in classical heavy-traffic analysis in the context of queueing networks, and also borrow from ideas in chemistry and biology. Such approaches have recently experienced something of a revival in the context of general massive interacting computational systems.

In this talk, we will introduce these approaches and showcase some of the methods and the results which can be obtained.

PDF of presentation slides (2.2 megabytes)
(need help viewing PDF files?)

Information from pubs.doc.ic.ac.uk/atnc-scalable-analysis.