Performance measurement of a shared-memory multiprocessor using hardware instrumentation

Abstract

A hardware approach is presented for the design of performance measurement insmentation for a shared-memory, tightly coupled MIMD multiprocessor. The Resource Measurement System (REMS) is a non-intrusive, hardware measurement tool used to obtain both trace measurement and resource utilization information. This approach provides more detailed and extensive measurement information than alternative software or hybrid approaches without introducing artifacts into the test results. This is accomplished at a significantly higher tool cost than the alternative software or hybrid approaches. Certain features of todays microprocessors limit the applicability of such a hardware tool. Measurements obtained using this hardware tool on two kemel (smaU benchmark) routines are presented.