Just-in-time dynamic voltage scaling: Exploiting inter-node slack to save energy in MPI programs

摘要: Although users of high-performance computing are most interested in raw performance, both energy and power consumption have become critical concerns. As a result, improving efficiency nodes on HPC machines has important, the prevalence power-scalable clusters, where frequency voltage can be dynamically modified, increased. On one opportunity for saving with little or no loss performance exists when computational load is not perfectly balanced. This situation occurs frequently, as keeping balanced between long-standing fundamental problems parallel distributed computing. Indeed, despite large body research aimed at balancing statically dynamically, this problem quite difficult to solve. paper presents system called Jitter that reduces assigned less computation and, therefore, idle slack time. saves these nodes, goal attempt ensure they arrive ''just time'' so avoid increasing overall execution Specifically, we determine which enough time such execute reduced cost-which will greatly reduce consumed node. In particular, 12.8% 0.4% increase-which essentially same hand-tuned solution-on Aztec benchmark.