The focus of our instrumentation work at NIST is to provide hardware support in obtaining performance measurement data of parallel computers, as well as uniprocessors, with tolerable perturbation to both the executing processes and the architecture on which they are executing. Tracing events and counting are two basic concepts of performance measurement-both should be controlled from within the code. The concept of tracing events is to follow selected points within execution paths so that execution times, delays, and event correlation can be computed. The concept of counting provides the basic mechanism for clocks, stop watches, histograms, etc., which can tally events at high frequencies-a must for high speed hardware events. Various forms of performance counters are starting to appear on commercial processor chips (e.g., Intel Pentium, MIPS R10000, IBM Power2, DEC Alpha). The exploration of instrumentation to provide these basic concepts in various ways within the constraints of low perturbation, low cost, and similar circuit technology has been the focus of our work
Published in:
Computer Performance and Dependability Symposium, 1996., Proceedings of IEEE International
Date of Conference: 4-6 Sep 1996