Performance Prediction for Data Intensive Applications on Large Scale Parallel Systems

摘要

This paper presents a new interactive performance estimation tool - PetaSIM for large scale parallel systems. Our main approach is to divide the difficult performance estimation problem into three domains: application, software and hardware, to extract the system specifications and provide tools for the interactive changes of the system parameters over the Internet. Computers, networks and applications are described as objects with special attention to the proper representation of caches and hierarchical memories. PetaSIM represents a prototype of a performance specification language (PSL). We present encouraging initial results for a set of data-intensive applications form the real applications and discuss the extension of PetaSIM to support applications in distributed collaborative engineering.