Non-intrusive and interactive profiling in parasight

Abstract

Debugging the performance of parallel applications is crucial for fully utilizing the potential of multiprocessor hardware. This paper describes profiling tools in Parasight, a programming environment that is geared towards non-intrusive performance analysis and high-level debugging. In Parasight, profilers execute as observer programs which run concurrently with the target program and monitor its behavior. This design grew out of our experience in debugging and monitoring the performance of real parallel programs. Parasight is used to implement an interactive profiler based on gprof, a standard UNIX profiling tool. The new profiler is an interactive tool that is dynamically controlled by the programmer and can run concurrently with the target program. It is also less intrusive than the original. This was achieved in part by off-loading the profile gathering computations to an adjacent processor and in part by allowing selective profiling. In conjunction with the new profiler, Parasight is used to implement high-level custom profiling routines that may be inserted at arbitrary locations in a target program to measure user-defined abstractions. One outcome has been to develop the beginnings of a profiling methodology in which high-level performance problems are detected through a function profiler and detailed performance bottlenecks are analyzed with custom profiling routines. Parasight, the new profiler and their associated tools have been implemented on the Encore Multimax, a shared-memory multiprocessor running UNIX. They have been used to profile both parallel and sequential programs.