Realistic CPU Workloads Through Host Load Trace Playback

Peter A. Dinda
pdinda@cs.nwu.edu
School of Computer Science, Carnegie Mellon University, and
Department of Computer Science, Northwestern University,

David R. O'Hallaron
droh@cs.cmu.edu
School of Computer Science and Department of Electrical and
Computer Engineering, Carnegie Mellon University

This paper introduces host load trace playback, a new technique for
generating a background workload from a trace of the Unix load average
that results in realistic and repeatable CPU contention behavior.
Such workloads are invaluable for evaluating various forms of
distributed middleware, including resource prediction systems and
application-level schedulers.  We describe the technique and then
evaluate a tool, playload, that implements it. Playload faithfully
reproduces workloads from traces on the four platforms on which we
have evaluated it.  Both playload and a large set of host load traces
are publicly available from the web at the following URL:
http://www.cs.cmu.edu/~pdinda/LoadTraces