high performance computing on graphics processing units: hgpu.org

With the increasing diversity of many-core processors, it becomes more and more difficult to guarantee performance portability with a unified programming model. The main reason lies in the architecture disparities, e.g., CPUs and GPUs have different architectural features from each other, which leads to the differences in performance optimization techniques. Thus, it is of great necessity to abstract performance-wise key features from many-core processors. In this paper, taking the Intel’s Xeon Phi as a case study, we present a two-stage comparative approach to abstract key features. Our approach needs a reference processor and is executed at both the application level and the microbenchmark level. We select multiple benchmarks from the Parboil benchmarks and measure the performance differences to identify performance factors. Further, we perform an in-depth analysis to identify the key features with microbenchmarks. Finally, we briefly discuss a use case in our optimizing framework.