RPPM : rapid performance prediction of multithreaded applications on multicore hardware

This paper proposes RPPM which, based on a microarchitecture-independent profile of a multithreaded application, predicts its performance on a previously unseen multicore platform. RPPM breaks up multithreaded program execution into epochs based on synchronization primitives, and then predicts per-epoch active execution times for each thread and synchronization overhead to arrive at a prediction for overall application performance. RPPM predicts performance within 12 percent on average (27 percent max error) compared to cycle-level simulation. We present a case study to illustrate that RPPM can be used for making accurate multicore design trade-offs early in the design cycle