The Design And Implementation Of A Framework Of Evaluation On HPC Platform

Now, HPC field need change from" creation machines without application requirement" to "creation machine for application ". To do this, the object and accurate benchmark suites is necessary which has become a hot point in HPC field.In this paper, a framework of evaluation for HPC platform is put forth which intend to evaluate the machines according to the requirement of application meet. The framework consists of three parts: machine profiling, application signaturing, evaluation modeling.Machine profiling bases on HPCC benchmark suite which can evaluate performance of machines in computation, communication and access memory. But there are two issues for users: one is identifying input data, the other is understanding test results. In this paper, an attempt is make to resolve the issues through auto-optimization of input file and post processing of output file. This will help to implement automatic HPCC.Application signaturing is divided two parts: single-processor performance, communication performance. For the former, hardware performance counters is used because it can accurately obtain performance events in low-level with minimal overhead of system. Compared to usual method applied to communication, the method used in the paper is relatively simple, but it can satisfy the need of our project.The evaluation modeling is built by forming the computational mapping of an application's signature onto machine profiles. Finally, we get the value of each machine according to capacity of meet requirement of application that is evaluation result.After validation of practical application, the result of the framework is incredible. The framework is able to form correct evaluation of HPC for application.