A Study on Optimal Scheduling Using High-Bandwidth Memory of Knights Landing Processor

In this paper, we profile three types of MPI workloads and propose an optimal scheduling scheme for Knights Landing (KNL), Intel's second-generation manycore processor, which contains high-bandwidth memory (HBM) with five times more bandwidth than existing DRAM. In addition, the KNL system can be configured for various forms of memory and cluster modes. To gain the best performance, it is important to analyze these systems' characteristics and schedule their workloads appropriately. By allocating memory-intensive workloads to HBM in KNL, we show that the overall performance of multiple MPI workloads can be improved in terms of execution time and system utilization.