Next: Performance and Scalability Up: Application of Massively Parallel Previous: Out-of-Core Algorithm

Performance and Numerical Results

In this chapter, we present performance measurement data for the parallel algorithms, which are given in the previous Chapter, implemented on three parallel platforms. Measuring the performance of the parallel implementation, we present the following two test procedures: 1) with a fixed machine size, we change the problem size to see the run time of all components of the implementation; 2) with a fixed problem size, we change the machine size to see the run time and speed-up for the dominant parts of the implementation. The scalability analysis is given for each parallel algorithm in the parallel implementations. Portability is demonstrated by porting the PVM implementation from Intel machines to IBM SP-1 with little work.

The numerical accuracy of the algorithms developed in Chapters 2 and 3 is investigated by running test cases provided by the Electromagnetic Code Consortium (EMCC). The results obtained by the ParaMoM-MPP code are presented for comparison with measurements provided by the EMCC done on their well-known benchmark targets. A discussion of the suitability of multiprocessing architectures for electromagnetic scattering problems is given. A conclusion drawn from this work is presented in this Chapter.

This chapter is divided into 3 sections: Section discusses the performance measurement and scalability analysis, Section provides the numerical results of the EMCC testing cases, and Section presents a discussion and comparison of the performance between the PATCH code and the ParaMoM-MPP code.



xshen@
Sat Dec 3 17:51:03 EST 1994