OAWS: Memory Occlusion Aware Warp Scheduling

We have closely examined GPU resource utilization when executing memory-intensive benchmarks. Our detailed analysis of GPU global memory accesses reveals that divergent loads can lead to the occlusion of Load-Store units, resulting in quick consumption of MSHR entries. Such memory occlusion prevents other ready memory instructions from accessing L1 data cache, eventually stalling warp schedulers and degrading the overall performance. We have designed memory Occlusion Aware Warp Scheduling (OAWS) that can dynamically predict the demand of MSHR entries of divergent memory instructions, and maximize the number of concurrent warps such that their aggregate MSHR consumptions are within the MSHR capacity. Our dynamic OAWS policy can prevent memory occlusions and effectively leverage more MSHR entries for better IPC performance for GPU. Experimental results show that the static and dynamic versions of OAWS achieve 36.7% and 73.1% performance improvement, compared to the baseline GTO scheduling. Particularly, dynamic OAWS outperforms MASCAR, CCWS, and SWL-Best by 70.1%, 57.8%, and 11.4%, respectively.