DescriptionThe memory subsystem is responsible for a large fraction of the energy consumed by compute nodes in High Performance Computing (HPC) systems. The rapid increase in the number of cores has been accompanied by a proportional increase in the DRAM capacity and bandwidth. Thus, the memory system consumes a significant amount of the power budget available to a compute node. There is a broad research effort focused on power management techniques using DRAM low-power modes. However, memory power management still presents many challenges towards Exascale. In this thesis, the potential of Dynamic Voltage and Frequency memory Scaling (DVFS) is studied considering the ability to select different frequencies for different memory channels. The approach adopted is based on tuning voltage and frequency dynamically to maximize the energy savings while maintaining performance degradation within tolerable limits. It was observed that HPC workloads rarely require maximum bandwidth, and the bandwidth demand placed by applications is spread over different channels. Also, HPC applications do not use all the bandwidth in a sustained manner, and they have phases where this bandwidth demand is not at its peak. Hence applications can tolerate reduction in bandwidth to save energy. Channel access patterns of applications are studied to determine the potential additional energy savings by controlling channels independently. Evaluation of proposed DVFS algorithm is conducted through a novel hybrid evaluation methodology that includes simulation and executions on real hardware. Results show the large potential of adaptive memory power management techniques based on DVFS for HPC workloads.