Evaluation of Impact of Noise on Collective Algorithms in Repeated Computation Cycle

スポンサーリンク

概要

• 論文の詳細を見る

• With the scale increasing, the impact of OS noise has become a key limiter of the performance of supercomputer. In this paper, we focus on a common supercomputing structure, which repeatedly execute computation followed by a collective operation implemented by using butterfly algorithm. We introduce a new proposed algorithm by increasing the number of paths of data exchange and evaluate the impact of noise on conventional butterfly algorithm and the proposed one by using the LogGOPS model. We simulate three situations: collective using butterfly algorithm without noise, collective using butterfly algorithm with noise and collective using proposed algorithm with noise. And the results show that the proposed algorithm improved the performance by almost 40 percent.

• 2013-09-23

著者

• Reiji Suda

  Presently With Crest Jst

• Reiji Suda

  Department o f Computer Science, Graduate School of In formation Science and Technology, University of Tokyo

• Hongzhi Chen

  Department o f Computer Science, Graduate School of In formation Science and Technology, University of Tokyo

関連論文

• An execution time prediction analytical model for GPU with instruction-level and thread-level parallelism awareness
• A precise measurement tool for power dissipation of CUDA kernels
• A Three-Step Performance Automatic Tuning Strategy using Statistical Model for OpenCL Implementation of Krylov Subspace Methods
• Efficient Monte Carlo Optimization with ATMathCoreLib
• Evaluation of Impact of Noise on Collective Algorithms in Repeated Computation Cycle
• The Future of Accelerator Programming: Abstraction, Performance or Can We Have Both?

スポンサーリンク