It would be highly desirable for operating systems to take a greater responsibility for process placement and load balancing decisions in parallel machines. This would relieve the programmer of much of the burden associated with fine-tuning an application in order to achieve acceptable performance levels. Before such operating systems can be developed, it is necessary to gain a better understanding of the factors that influence program performance. In particular, it would be useful to be able to identify <I>classes</I> of programs which, in a statistical sense, behaved in a similar manner. Then, given an arbitrary program whose class was known, rules and heuristics developed for the program class (in conjunction with program specific information) could be used to make informed placement and load balancing decisions. As a step in this direction, this thesis investigates the application of standard statistical techniques to the performance analysis of particular classes of parallel programs. Simple CSP-type parallel programs exhibiting loosely synchronous data parallelism are used to illustrate how a common class of programs can be characterised in terms of a relatively small number of parameters representing time-averaged properties. In order to systematically explore parameter space, synthetic programs are used. The execution of these programs is simulated on an accurate performance model of a transputer-based machine. Standard experimental design techniques, such as the analysis of variance, are then applied to develop statistical models relating to the program class. It is shown that useful quantitative predictions can be made for arbitrary class members.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:660590 |
| Date | January 1994 |
| Creators | Phillips, Joseph |
| Publisher | University of Edinburgh |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://hdl.handle.net/1842/15648 |
Page generated in 0.0021 seconds