Energy Efficient Parallel Matrix-Matrix Multiplication for DVFS-enabled Clusters

Excessive energy consumption has become one of the major challenges in high performance computing. Reducing the energy consumption of frequently used high performance computing applications not only saves the energy cost but also reduces the greenhouse gas emissions. This paper focuses on developing energy efficient algorithms and software for the widely used matrix-matrix multiplication, so that it is able to consume less energy in a DVFS-enabled cluster with little sacrifice in performance. The state-of-the-art practical parallel matrix matrix multiplication algorithm in ScaLAPACK partitions matrices into small blocks and distributes matrices using a two dimensional block cyclic distribution approach. Experimental results demonstrate that our energy efficient matrix-matrix multiplication algorithm can save up to 26.35% of energy with about 1% performance penalty. And the modified PDGEMM of ScaLAPACK is able to save energy more than 20% with less than 2% of performance loss.