Performance Modeling And Prediction Of HPCCG

For both software developers and software users, the establishment of softwareperformance model has important significance. Using the performance models,software developers can have a deeper understanding of the software, so they candevelop better and more efficient software; software users can make full use of thesoftware and can supervise the software more easily. However, with the increase ofpeople’s needs, the software has become increasingly large scale, coupled with thedevelopment of computer hardware technology, computer architecture becomesvery complex. Building an accurate performance model flexibly and efficientlybecomes growingly difficult.This paper proposes three analytical methods to establish the performancemodel including the problem size estimation method, floating-point estimationapproach and integration by parts method. HPCCG is a small benchmark developedby the Sandia National Laboratories. Its kernel algorithm is conjugate gradientsolver running on a lot of high-performance computers. This paper chooses HPCCGas the research object to build its performance model using the methods above, thentakes advantage of the performance models to predicate runtime of the software andanalyses the error.