Page management in hybrid memory systems
Apostolos Kokolis · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2019
Recent byte-addressable Non-Volatile Memory (NVM) technologies enable hybrid memory systems comprising of both DRAM and NVM technologies. Such systems have the potential to address the capacity requirements of data intensive workloads and achieve high performance. The main challenge lies in dynamically managing data placement between DRAM and NVM in a flat address space configuration, where the Operating System can allocate pages to either of the two memories. Prior work on this area has proposed software and architectural techniques that can dynamically swap memory pages between the two memories. However, due to the high swap overhead, initiating swaps solely based on hardware counters or relying on software methods hinders the potential for performance gains due to their conservative decision making. In this work, we introduce Prefetching, a novel hybrid memory management scheme that exploits page correlation to identify forthcoming memory accesses and commence swaps ahead of time. The page management techniques proposed by Prefetching effectively hide the overhead of data movement between memories. We evaluate our design with simulations across 17 benchmarks from three different benchmark suites. Thanks to Prefetching's highly accurate swaps, we improve performance by up to 20% and reduce average main memory access time by up to 33% when compared to prior state-of-the-art.