Concurrent programming using Parallel Virtual Machine.

摘要

The thesis deals with the design, development and performance evaluation of concurrent programs using Parallel Virtual Machine (PVM) software system. We consider Matrix Multiplication, Sorting, Embarrassingly Parallel Kernel and Multigrid Kernel and implement their concurrent programs in PVM on an IBM SP2 machine. The structure of PVM programs and different types of parallel paradigms such as master-slave, farming and ring for the selected applications are discussed. Strategies such as the use of barriers and groups are incorporated in the implementations to achieve synchronization of processes.;We also focus on issues including the ease of development of parallel code, efficiency of remote machines with respect to ease of use, computing times, response times and other issues which affect the parallel programmer. Experiences of a remote user of a parallel computer are given to exemplify the challenges involved in using a remote machine.;For the concurrent implementations, speedups ranging between 1.0-7.7 were observed for 2-8 processors. Poor speedups area attributed to the high intercommunication costs whereas the absence of any significant amount of inter-process communication and concurrcent communications between the processes account for the high speedups. Performance models are developed to describe the timings obtained on an IBM SP2. The models can be used to predict the asymptotic values of speedup and efficiency of the concurrent implementations.;The PVM programs can be implemented on other parallel machines and a wide range of problems can be parallelized using PVM. The performance models presented can be adapted to other IBM SP2 installations as well as other parallel machines.