Performance Evaluation of Hybrid Parallel Programming Paradigms

摘要: With the trend in supercomputing world shifting from homogeneous machine architectures to hybrid clusters of SMP nodes, interoperabiility OpenMP and MPI has become a key issue understanding optimizing overall system performance. While low-level performance can be evaluated using existing benchmarks, combination two poses new challenges. Therefore, study different programming paradigms is high benefit for both vendors user community. As part our project, we have identified several possible combinations models order provide qualitative quantitative justification situations which any one them favoured. Collective operations are particularly important analyze evaluate on platform therefore concentrate three -- barrier, all-to-all, all-reduce. Issues like optimal mix MPI, most efficient way managing communication within OpenMP, unit communication, degree overlap between computation need evaluated. The results supporting this investigation were taken IBM Power-3 at San Diego Supecomputer Center suite microbenchmarks.