A quantitative evaluation of unified memory in GPUs

摘要

The introduction of unified memory and demand paging has simplified programming of graphics processing units (GPUs). It has also enabled oversubscribing the memory for a GPU. However, the overhead of page management makes page faults a performance bottleneck. Sometimes the page eviction policy is unable to mitigate performance slowdown caused by page faults and memory oversubscription. On average, eviction policies such as Random and CAR are not competitive with a traditional least recently used (LRU) policy. Other policies, such as CLOCK-Pro, are designed to overcome challenges with LRU, but they only achieve limited speedup. Even enhancing LRU with page walk hit information does not lead to notable performance improvement. Based on these observations, we propose optimization opportunities to mitigate performance degradation caused by page faults and memory oversubscription. These optimization opportunities include an effective page eviction policy that retains LRU's advantages while addressing LRU's inability to deal with thrashing access patterns, page prefetch and pre-eviction, memory-aware throttling, and capacity compression.