A multiprocessor execution profiler

The authors have examined existing profiling tools, which generally have crude interfaces, are clumsy to use, and monitor only accumulated CPU time and function calls, and have concluded that they are inadequate for profiling a large-scale multiprocessing system, even if they are adequate for manual analysis of a single processor system. They have developed a tool that collects the following information: where CPU time is expended, quality and quantity data passed between functions, how much global data is referenced and modified, and how these characteristics differ among processors on the network. To simplify data inspection the profiler has a mouse-driven graphical interface. The authors describe the design of the profiler and give examples of its utilization.<>