Selective Memory Compression for GPU Memory Oversubscription Management

Unified virtual memory for discrete GPUs helps increase programmer productivity by abstracting the presence of different memories. However, fault-driven migrations and remote fault handling slow down the applications considerably. During memory oversubscription, the overheads increase drastically due to increased page thrashing and page evictions. We propose a GPU memory compression system, Selective Memory Compression (SMC), that selectively compresses read-only pages to increase the effective memory size while avoiding costly page remappings due to page overflows. We also propose a line packing scheme for compressed pages, Split Linearly Compressed Pages (SLCP), that minimizes unused space, gives better compressibility, and reduces extra memory accesses performed to fetch data in a compressed memory system. We show that under 125% and 150% oversubscription, SMC combined with SLCP gives 53% and 60% performance improvement, respectively, over a baseline that uses the state-of-the-art eviction policy.