Using Recorded Values for Bounding the Minimum Completion Time in Multiprocessors
IEEE Transactions on Parallel and Distributed Systems
VPPB -- A Visualization and Performance Prediction Tool for Multithreaded Solaris Programs
IPPS '98 Proceedings of the 12th. International Parallel Processing Symposium on International Parallel Processing Symposium
Hi-index | 0.00 |
In this paper we present a method an a set of tools forpredicting the speedup of multithreaded Solarisprograms. The predictions are based on recordings froma single-processor execution of the multithreadedprogram. The routines in the thread library areoverloaded with an instrumented thread librarydeveloped by us. We do not need to have access to thesource code of the multithreaded program and norecompilation is needed.Using three multithreaded programs and amultiprocessor with 8 processors, we are able to showthat the predictions are very reliable, particularly forprograms with a coarse grained parallelism. The run-timeoverhead for the recordings is very small. Thepresented technique does not model I/O, and is thereforeapplicable only to CPU-intensive applications.