Using Processor Partitioning to Evaluate the Performance of MPI, OpenMP and Hybrid Parallel Applications on Dual- and Quad-core Cray XT4 Systems
Xingfu Wu, Valerie Taylor · 2009
Abstract: Chip multiprocessors (CMP) are widely used for high performance computing. While this presents significant new opportunities, such as on-chip high inter-core bandwidth and low inter-core latency, it also presents new challenges in the form of inter-core resource conflict and contention. A challenge to be addressed is how well current parallel programming paradigms, such as MPI, OpenMP and hybrid, exploit the potential offered by such CMP clusters for scientific applications. In this paper, we use processor partitioning as a term about how many cores per node to use for application execution to analyze and compare the performance of MPI, OpenMP and hybrid parallel applications on two dual- and quad-core Cray XT4 systems, Jaguar with quad-core at Oak Ridge National Laboratory (ORNL) and Franklin with dual-core at the DOE National Energy Research Scientific Computing Center (NERSC). We conduct detailed performance experiments to identify the major application characteristics that affect processor partitioning. The experimental results indicate that processor partitioning can have a significant impact on performance of a parallel scientific application as determined by its communication and memory requirements. We also use the STREAM memory benchmarks and Intel’s MPI benchmarks to explore the performance impact of different application characteristics. The results are then utilized to explain the performance results of processor partitioning using NAS Parallel Benchmarks. In addition to using these benchmarks, we also use a flagship SciDAC fusion microturbulence code (hybrid MPI/OpenMP): a 3D particle-in-cell application Gyrokinetic Toroidal Code (GTC) in magnetic fusion to analyze and compare the performance of these MPI and hybrid programs on the dual- and quad-core Cray XT4 systems, and study their scalability on up to 8192 cores. Based on the performance of GTC on up to 8192 cores, we use Prophesy system to online generate its performance models to predict its performance on more than 10,000 cores on the two Cray XT4 systems.