Power-aware numerical methods on heterogeneous grids (コンピュータシステム)

概要

論文の詳細を見る
Recent high performance computing systems have pursued a double objective of performance and low-power consumption. Related researches have mainly focused on low-power architectures and power-aware computing by means of Dynamic Voltage Scaling (DVS). Our work aims to extend power-aware constraints to heterogeneous grid computing for the resolution of numerical problems. We have first focused on the eigenproblem using the explicit restarted Lanczos method. We have stressed the great impact of the CPU frequency heterogeneity of computing resources and, we have shown how DVS allows to decrease the global (resp. local) energy consumption by a factor of 9% (reap. 20%) without any significant increase of the wall-clock time. Ongoing research concerns the hybrid GMRES/LS-Arnoldi method and multigrid numerical methods. The hybrid GMRES/LS-Arnoldi runs in parallel 3 different numerical methods and uses asynchronous communications between them. We believe this pattern has a potential for significant energy saving. The parallel multigrid method takes advantage of the parametric parallelism. Besides, with the domain decomposition, the intra-node grid levels and the grain of grids, we have multiple levers for using DVS.
社団法人電子情報通信学会の論文
2008-07-29