http://worldcat.org/entity/work/id/355527303

The implementation and evaluation of a parallel algorithm on C. MMP

The purpose of this paper is to get a detailed look at the performance of an implementation of a parallel program on C.mmp. The rootfinding algorithm was chosen because it meets two constraints: it is a parallel algorithm with significant interprocess communication; and it is of relatively low complexity, enabling us to focus on implementation issues rather than subtleties in the algorithm itself. Variations in processor speeds and asynchronously executing operating system functions are shown to have a detrimental effect on the rootfinder's performance. However, the most important implementation decision affecting the performance of the rootfinding program is the type of synchronization semaphore used. We define the threshold for practical application of a semaphore to be when 50% of the execution time is attributed to semaphore related overheads. Using the 50% criteria, we measured thresholds for inter-synchronization times from two milliseconds for the most primitive locks, to 200 milliseconds for the most sophisticated and flexible semaphore. During the course of these measurements, Hydra underwent several revisions and the 200 millisecond threshold was reduced to 33 milliseconds. The principal concept responsible for this performance improvement is discussed in the paper.

Open All Close All