A Run-Time Performance Monitor for Message-Passing Parallel Programs

A fundamental problem with parallel program monitoring tools is the intrusiveness introduced by instrumenting a monitor into the original program for collecting trace data. An intrusive monitor may alter the order of program events and distort the timing information. This may result in a change in the probability of making particular non-deterministic choices, and thus change the program behaviour. This paper presents an instrumentation approach for monitoring parallel message-passing programs with a minimum intrusiveness. The approach uses virtual clocks to time and control the ordering of communication events during monitoring, and to reflect the real time execution when running without monitoring. Its main improvement over some previous approaches is that it not only preserves the partial ordering of events, but also collecting accurate timing information. It is therefore suitable for both program debugging and performance tuning. The paper describes how the approach collects timing..