Invasiveness of Performance Instrumentation Measurements on Mulitprocessors

Performance is a critical issue in order to justify the use of parallel computers. Since it is usually a difficult task to write an application that successfully exploits the target parallel architecture, many tools like Pablo, ParaGraph, PA-Tools, Express, TOPSYS and ANDES have been developed. These tools use different software, hardware or hybrid mechanisms to record events of interest related to performance. Nevertheless, in a multiprocessor environment, the degree to which an event collection mechanism perturbs the behaviour of the application being monitored may become a significant factor. This factor, often called invasiveness, reduces the accuracy of the performance data and may lead the user to wrong conclusions. For this reason, invasiveness must be taken into account when measuring the performance of an application. The invasiveness of the trace collection mechanism of PARMACS, a set of macros which provides a portable programming model, is determined on four different hardware platforms and for two different applications (including an LU Matrix Decomposition Algorithm). The purpose of this paper is to determine the degree of invasiveness of PARMACS and how this invasiveness depends on the application, instrumentation and hardware platform. These results are compared to similar tools and a state of the art survey in this field is also presented. Keyword Codes: D.2.8; D.4.8; D.2.2