Parallel processing of recurrence operations in SMP machines

Abstract

The parallel processing techniques have been widely researched in an effort to improve the performance of processing large and complex operations in both numerical and non-numerical applications. One of the typical approach is the automatic parallelizing compilers which transform sequential programs to parallel programs. However, many scientific operations, such as recurrence operation, are not yet suitably treated by parallel processing because of their intrinsic difficulty of parallelizing and peculiar attributes in their formula despite they are important and widely used. There are several researches on the first-order linear recurrence operation using vector super=computers achieving considerable performance enhancement. In this thesis, we focus on parallelization of recurrence operations in SMP machines. Recently , SMP machines have rapidly been developed and their performance and scalability have shown considerable progress. In the present, SUN and Intel are supplying the server market with various models of SMP machines, such as 4-way, 8-way, and 64-way system. However, there is few work related to parallel processing of scientific computation using SMP machines. Recurrence operation has typical characteristic which are required in many basic scientific computations. So the research on the parallel processing of recurrence operations in SMP machines can be generally applied to any other scientific and mathematical computations. We have employed various algorithms using thread technique in one-dimensional and two-dimensional recurrence operations on SMP machines. We have evaluated and analyzed the performance results. We have observed reasonable performance enhancement. The memory contention and some problems related to memory architectures predicted in the beginning of the work turned out to be to longer serious problems in the small-scale SMP machines. Cache memory has a great role in the performance of SMP machines even in scientific computations.