A Cost-Based Approach to Performance Analysis of Distributed Object
Stores
S. P. Fenwick
Department of Computer Science,
The Australian National University,
Canberra, ACT 0200, Australia.
stevef@cs.anu.edu.au
C. W. Johnson
Department of Computer Science,
The Australian National University,
Canberra, ACT 0200, Australia.
cwj@cs.anu.edu.au
Abstract
Monitoring the performance of programs running on massively parallel
computers or distributed systems is an extremely difficult problem
because of the volume of information generated and the complexity of
interactions between components of the program. This problem is made
even more difficult by the new parallel programming paradigm provided
by ``distributed object stores''. This paper discusses the problems
of analysing the performance of programs that use distributed object
stores. A conceptual analysis framework that addresses many of these
problems is then presented. This framework models the object store as
a collection of resources, each of which has a state that can vary
during execution. Operations performed by the program are modelled as
requests to these resources, the cost of which can depend on the
resource's state. The performance of the program can then be analysed
in terms of the states and costs of the object store resources that it
accesses, providing a means of detecting various forms of contention
between client processes. This approach allows the programmer of a
system that uses a distributed object store to analyse performance
problems without needing a detailed understanding of the
implementation of the object store.
Conference Home Page