Performance visualization of parallel programs on a shared memory multiprocessor system

Abstract

A performance-monitoring tool for visualizing the execution of parallel programs in a shared memory multiprocessor system is presented. A parallel program that contains pieces of work can be described, according to the parallelizing constructs in the program, as program sections, such as serial and parallel sections. The tool displays on a section-by-section basis the utilization of the participating processes according to where their execution time is spent: user and system work, parallelization overhead, idle time, etc. The performance data can be visualized during the execution of the program or can be recorded in event traces and displayed to the user after the parallel program is terminated. Taking advantage of the X window system and high-resolution graphic displays, the tool provides several color plots for understanding how efficiently an application program is executed on a shared-memory multiprocessor system. In particular, it guides a user to locate any bottlenecks which may prohibit a parallel program from utilizing the system.