OpenMP benchmark using PARKBENCH

Real application codes in OpenMP obviously measure the performance of OpenMP programming on the real problems. Although this is ultimately what the end-user wants, the full real applications are often complex and large. In order to obtain a guide to the performance of OpenMP parallel programs in any given parallel systems, kernel and synthetic benchmarks are useful. PARKBENCH[4] is a set of benchmark programs proposed by PARKBENCH committee. It consists of synthetic low-level benchmarks, kernel benchmarks and compact applications. This hierarchical structure allows information derived from the simpler code to be used in explaining the performance characteristics of the more complicated codes. Currently, the low-level and kernel benchmarks are available. The kernel benchmarks includes some kernel programs from the NAS parallel benchmarks. In PARKBENCH, fundamental metrics such as the accuracy of clock and the op count is clearly de ned. In the current version, the parallel benchmarks are written using MPI and PVM to evaluate parallel program in distributed memory environments. In this paper, we focus on the single processor benchmarks in the low-level benchmark of PARKBENCH as an OpenMP benchmarks. We have parallelized these benchmarks using OpenMP. To understand the performance of OpenMP programs, it is important to measure the performance of shared memory hardware platforms as well as the OpenMP constructs overheads. The benchmark programs in OpenMP shows the performance of a set of loops parallelized by using OpenMP programming model. The benchmark assesses the signi cance of the overheads of OpenMP loop constructs in several kinds of loops. The overheads of the parallel constructs should be evaluated with respect to shared memory performance. This benchmark results show a guide to parallelize loops using OpenMP. For example, the OpenMP programmer can identify the e ective loop-length in OpenMP for several kinds of loops. Some OpenMP benchmarks were already proposed. EPCC OpenMPmicrobenchmark[2] is a synthetic benchmark to measure the overheads of synchronization primitives and loop scheduling for language constructs in OpenMP. Although this benchmark gives a valuable information about the overheads of OpenMP constructs, it does not address the performance of the shared memory hardware. The NAS parallel benchmark suite[3] is the most popular benchmark for distributed memory platforms. Recently, a OpenMP version of the benchmarks were already prepared, and is expected to be available soon. As realistic application benchmarks, One of SPEChpc96, SPECclimate is parallelized by OpenMP. SPEC HPG [6] has announced a plan of SPEC OpenMP benchmarks as one of SPEC benchmark suites for high performance computing eld. In the next section, we introduce the PARKBENCH benchmarks brie y and the OpenMP version of its low-level benchmark. Section 3 presents the results of the measurements on the di erent systems and di erent compilers, and the analysis by PARKBENCH outputs. We conclude our work in Section 4.