WHAT DO WE GAIN FROM HYPER-SYSTOLIC ALGORITHMS ON CLUSTER COMPUTERS?

摘要

The principle problem of parallel computing, the degradation of performance by internode-communication, is aggravated on modular cluster-architectures with off-board communication cards. In this contribution we consider so-called N-squared problems, the prototype being the N-body computation with long range interaction. We demonstrate that hyper-systolic algorithms are able to enhance the performance of such computations on parallel machines by alleviating the communication bottleneck, at the cost of moderately increased memory. We determine the efficiency of implementations on the Alpha-Linux-Cluster-Engine ALiCE at Wuppertal university compared to the SIMD system APE100.