Performance Characterization and Optimization for Intel Xeon Phi Coprocessor

Abstract

The Intel Xeon Phi is a many-core accelerator which focuses on the high performance applications. To characterize the performance of the Intel Xeon Phi, a system of dual 8-core Intel Xeon E5-2670 processors is employed as a control platform, and a subset of the PARSEC benchmark suite is selected as the benchmark applications. The first evaluation in this paper shows that the applications on the Intel Xeon Phi is averagely 2.06x slower than on the dual Intel Xeon E5-2670. The further detailed performance characterization quantifies the performance impact of various architecture parameters on the Intel Xeon Phi. To set an example for how to improve the architecture of the Intel Xeon Phi for better performance, the hardware optimization with an additional set of vector processing units is discussed and a simple emulator is developed accordingly. The evaluation results show that this optimization can provide an average speedup of 1.10.