Abstract:

Many recent data center servers are built with NUMA (Non-Uniform Memory Access) characteristics. Accessing remote memory generally takes longer time than accessing local memory. There are a lot of research works that discuss the performance improvement of NUMA multi-core systems. However, rare research work considers reducing the energy cost of NUMA multi-core systems. This work studies reducing energy cost of multi-threaded programs on NUMA architectures using DVFS (Dynamic Voltage and Frequency Scaling) adjustment strategy. We consider three factors of the multi-threaded programs which influence the energy saved by our DVFS adjustment strategy. These three factors are: (1) the memory access intensity of parallel programs; (2) the proportion of remote memory access; (3) the ratio between remote and local memory access latency. In addition, we propose two DVFS adjustment strategies to save the energy cost of multi-threaded programs. The energy-saving effect of these two DVFS adjustment strategies is influenced by these three factors. Two DVFS adjustment strategies can save maximally 20% and 39.2% of total energy when considering one factor and 33.3%, 48.1% of total energy when considering two factors, respectively.