Report ID
1996-13
Report Authors
Bjoern Haake
Report Date
Abstract
Because parallel programs are very complex, their execution is hard tounderstand. A tool that supports the analysis of a parallel program is calleda profiler. In this thesis, we compared different profiling techniques andexamined existing profiling tools. We used these informations to design andimplement a profiler for Split-C. Split-C is a parallel extension to C. Mostof the existing implementations are based on active messages, a fastcommunication mechanism. Split-C provides a global address space, where everyprocessor has access to. Split-phase assignments help to overlap computationwith communication. We use a version of Split-C that is implemented on theMeiko CS-2 parallel machine. The Meiko is built from SuperSparc processorsthat are interconnected via a fat-tree network. This resembles a network ofSparc 10 workstations. The Split-C profiler measures the differentcommunication routines and creates entries for a tracefile in the PICL format.The overhead introduced by the PICL calls varies, in most of the routines itranges from 1--10 \\%, only some of the routines imposed a larger overheadbetween 20 and 40 percent. Roundtrip communication adds larger overhead. Theoverhead is greater in Split-C compared to MPI (Message Passing Interface, theoriginal target for PICL) because Split-C communication routines areapproximately 10 times faster than MPI\'s. The tracefile can be postprocessedvia a graphical interface, such as ParaGraph. We used ParaGraph because it hasa very convenient display and uses colors for a helpful and understandableoutput. We achieved to use the Split-C profiler for improving some largeapplications.
Document
1996-13.ps6.29 MB