Risk-Theory Based Performance Analysis for Scale-Out Parallel Computing Systems

2014 Fiscal Year Final Research Report

• Project/Area Number: 25540007
• Research Category: Grant-in-Aid for Challenging Exploratory Research
• Allocation Type: Multi-year Fund
• Research Field: Mathematical informatics
• Research Institution: Nara Institute of Science and Technology
• Principal Investigator: KASAHARA Shoji 奈良先端科学技術大学院大学, 情報科学研究科, 教授 (20263139)
• Co-Investigator(Renkei-kenkyūsha): MASUYAMA Hiroyuki 京都大学, 大学院情報学研究科, 准教授 (60378833) ; TACHIBANA Takuji 福井大学, 大学院工学研究科, 准教授 (20415847)
• Project Period (FY): 2013-04-01 – 2015-03-31
• Keywords: 情報基礎 / クラウド・コンピューティング / 極値理論 / タスク・スケジューリング / 性能解析

Outline of Final Research Achievements

This study considered the risk-theory-based analysis of task scheduling for large-scale parallel distributed computing. It was found that in task replication, the optimal number of replications which achieves the shortest task-processing time mainly depends on the coefficient of variation of the worker-processing time, and that three replications are sufficient to guarantee a low variance of the task-processing time. In terms of backup-task scheduling, we analyzed system performance by extreme value theory. It was found that a small deadline time is effective for reducing energy consumption when the subtask-processing time follows a heavy-tailed distribution. In case of light-tailed subtask-processing time, on the other hand, energy consumption can be reduced by a large deadline time.

Free Research Field

待ち行列理論