• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2013, Vol. 35 ›› Issue (3): 25-30.

• 论文 • Previous Articles     Next Articles

Performance evaluation methodology for
massively parallel computer systems

LIU Jie,CHI Lihua,JIANG Jie,XU Han,YAN Yihui,HU Qingfeng   

  1. (National Laboratory for Parallel and Distributed Processing,Changsha 410073,China)
  • Received:2011-05-16 Revised:2011-08-20 Online:2013-03-25 Published:2013-03-25

Abstract:

The complexity of modern massively parallel computer systems requires sophisticated method to facilitate the evaluation of system performance, and it is difficult to measure all the characteristics of systems using a single metric. The concept of applicability is given from the demand of users for supercomputers. Based on the applicability concept, we present a performance evaluation methodology for massively parallel computer systems. The methodology, including total applicability, subentry applicability, the concepts of subentry, metrics and benchmarks, can avoid the personal factors and obtains the impersonality results for systems performance. After collecting data using performance analysis tools, the data can be used to get the subentry applicability. Then the total applicability can be calculated by the weight equation. Based on the total applicability, one can evaluate whether the massively parallel computer systems satisfy the demand of user or not.

Key words: performance evaluation;massively parallel computer system;applicability of user