Learning Based Performance and Power Efficient Cluster Resource Manager for CPU-GPU Cluster

摘要

The recent success in building petascale High Performance Computing (HPC) systems have produced the demand for efficient and optimized use of resources to increase the performance and reduce the power consumption. Including the above, the heterogeneous architectures of nowadays HPCs comprising a multicore CPU and many-core Accelerator like GPU(s) are facing another concern for using optimum utilization of each of these components. This paper presents the scheduling mechanism of the Cluster Resource Manager (CRM): i. Moldable job Scheduler (MS) which is able to mold the jobs with respect to the number of machines based on an preliminary initialized and auto updated heuristic knowledge-base of problem size, optimum machine count, execution duration to increase the utilization of the full cluster facility. ii) Collocation Aware and Power Efficient Resource Manager (CAPE-RM) manages collocation of CPU only and GPU accelerated jobs by monitoring the CPU load and memory usage. The emerging computation ability is followed by the huge amount of power consumption. Though the use of GPU(s) itself cut down the power to be needed by the only CPU based cluster but to make a green computing facility more power efficiency is desired. The CAPE-RM is designed to support the above by powering off the idle nodes by monitoring the total load to the facility and based on a simple statistic of the frequency of job submission.