Runtime Power Allocation Based on Multi-GPU Utilization in GAMESS

To improve the power consumption of parallel applications at runtime, modern processors provide frequency scaling and limiting capabilities. In this work, a runtime strategy is proposed to maximize performance under given budget by distributing available according relative GPU utilization. Time series forecasting methods were used develop workload prediction models that accurate utilization during application execution. Experiments performed on multi-GPU computing platform DGX-1 equipped with eight NVIDIA V100 GPUs for quantum chemistry calculations in GAMESS package. For limited budget, may deliver as much hundred times better than obtained when distributed equally among all GPUs.