Performance Analysis of Compiler-Parallelized Programs on Shared-Memory Multiprocessors

Shared-memory multiprocessor (SMP) machines have become widely available. As the user community grows, so does the importance of compilers that can translate standard, sequential programs onto this machine class. Substantial research has been done to develop sophisticated parallelization techniques, which can detect and exploit parallelism in many real applications. However, the performance of compiler-parallelized applications can be below expectations. The speedups of even fully-parallel codes on today's shared-memory multiprocessors can be signi cantly less than the number of processors. In this paper we will investigate reasons for such performance behavior. We will focus on three speci c issues: (1) We will determine whether it is appropriate for the preprocessor to express the detected parallelism in the common loop-oriented form, (2) we will determine sources of ine ciencies in fully parallel SMP programs that exhibit good cache locality, and (3) we will discuss the portability of these programs across SMP machines. In our experiments we have extended the Polaris compiler, so that it can generate threadbased code directly. We compare the performance of this code with Polaris' loop-parallel OpenMP output form and with architecture-speci c directive languages available on the Sun Enterprise and the SGI Origin systems. We have analyzed in detail the performance of several parallel Perfect Benchmarks. Our main ndings are as follows. (1) Overall, there is no signi cant performance disadvantage of the loop-parallel representation. (2) However, substantial performance di erences are attributable to the instruction e ciency, which is in uenced by the data sharing semantics of parallel constructs. (3) Both the OpenMP and the thread-based program forms are functionally portable, but can result in substantially di erent performance on the two machines. y This work was supported in part by DARPA contract #DABT63-95-C-0097 and NSF grants #9703180-CCR and #9872516-EIA. This work is not necessarily representative of the positions or policies of the U. S. Government.