Automatic Construction of Performance Skeletons for Grid Resource Selection

University Of Houston, Houston TX

Investigators

Abstract

National Science Foundation Distributed Systems Research CISE/CNS ABSTRACT PROPOSAL NUMBER: 0410797 PRINCIPAL INVESTIGATOR: Subhlok, Jaspal. INSTITUTION: University of Houston PROPOSAL TITLE: Automatic Performance Skeletons for Grid Computing The performance skeleton of an application is a short running program whose execution time in any run-time scenario reflects the expected execution time of the application it represents. Such a skeleton can be employed to quickly esti-mate the performance of a long running application on a given set of execution nodes even when the nodes and network links are shared. This project is per-forming research to enable automatic construction of performance skeletons of resource intensive applications. A performance skeleton must mimic application execution behavior, such as computation, memory, and communication pat-terns, in order to model the application's performance. The approach being taken to construct the performance skeleton of an application consists of the following steps: (1) Execute the target application and record the execution trace that captures detailed CPU, memory and network activity. (2) Capture repeated patterns in the application execution trace and summarize it as a compact execution signa-ture. (3) Generate the performance skeleton executable program from the summarized system activity captured by the execution signature. Performance modeling is central to effective use of high performance computa-tional resources. Performance skeletons represent a fast, innovative and accu-rate approach to estimating the performance of long running applications on diverse and unpredictable grid environments. Accurate performance estimation with this method typically takes only a few seconds of skeleton execution for applications that run for several hours. This project is also developing a frame-work for resource selection in grid environments with performance skeletons.