Quantifying the NUMA Behavior of Partitioned GPGPU Applications
Alexander Matz, Holger Fröning · 2019
While GPU Computing is pervasive in various areas, including scientific-technical computing and machine learning, single GPUs are often insufficient to meet application demand. Furthermore, multi-GPU processing is a promising option to achieve a continuing performance scaling, given that CMOS technology is expected to hit fundamental scaling limits. We observe that a large amount of work has been done regarding characterizing single-GPU applications. However, characterizing GPGPU applications regarding the NUMA effects resulting from distributed execution has been largely overlooked.