Performance Modeling of HPC Applications

Performance models of applications enable HPC system designers and centers to gain insight into the most optimal hardware for their applications, giving them valuable information into the components of hardware (for example processors or network) that for a certain investment of time/money will give the most bene t for the applications slated to run on the new system. The task of developing accurate performance models for scienti c application on such complex systems can be di cult. In section 2 we brie y review a framework we developed [1] that provides an automated means for carrying out performance modeling investigations. In section 3 we describe ongoing work to lower the overhead required for obtaining application signatures and also how we increased the level-of-detail of our convolutions with resulting improvements in modeling accuracy. In section 4 we show how these technology advances have enabled performance studies to explain why performance of applications such as POP (Parallel Ocean Program) [2], NLOM (Navy Layered Ocean Model) [3], and Cobalt60 [4] vary on di erent machines and to quantify the performance e ect of various components of the machines. In section 5 we generalize these results to show how these application's performance would likely improve if the underlying target machines were improved in various dimensions (as for example on future architectures).