high performance computing on graphics processing units: hgpu.org

As we move towards exascale computing, the efficiency of application performance and energy utilization, must be optimized by redefining architectural features and application performance analysis. This research analyzes the performance per core of 8 applications on Intel Xeon Phi Knights Corner (KNC) and Knights Landing (KNL) to determine if performance variation within cores can lead to performance and energy improvements. Our results showed that KNC architecture’s core vary in performance, leading to faster inner core performance as a result of memory characteristics and core utilization. It also shows that cores 17, 34, and 51 on the KNL architectures performs consistently slower than other cores, with core 0 performing either faster, slower or within the average performance time all the cores. A power performance study was then done utilizing different core configurations on the KNC. The results show that by targeting inner cores for applications that exhibit better inner core performance, a maximum energy reduction of 16.4% compared to a con- figuration using all cores was possible with its optimal thread configuration. Energy reduction was achieved with along with a 2% reduction in the fastest execution time of the same application. Our results also show how application characteristics lead to different core variation performances on KNC and KNL Xeon Phi architectures.