Impact of L2 cache locking on GPU performance

Cache locking is a proven method for increasing time predictability of general-purpose CPUs. In the Graphic Processing Unit (GPU) architecture, caches are smaller and serve a different overall purpose than their CPU counterparts; consequently, it needs to be determined if GPUs will benefit from cache locking. To examine the impact of cache locking on GPUs, we extended the functionality of GPGPU-Sim, a GPU simulator, to support L2 cache locking and randomly selected six benchmarks from the Rodinia benchmark suite to serve as the datasets. In our experiment, both the cache size and cache associativity were varied and the most frequently accessed data were locked into the L2 cache. The overall results obtained show that locking data into the GPU´s L2 cache does not degrade the GPU performance significantly while improving time predictability. Also, for some benchmark, L2 cache locking actually improves the GPU performance by up to 56%.