ATMem: adaptive data placement in graph applications on heterogeneous memories

Active development in new memory devices, such as non-volatile memories and high-bandwidth memories, brings heterogeneous memory systems (HMS) as a promising solution for implementing large-scale memory systems with cost, area, and power limitations. Typical HMS consists of a small-capacity high-performance memory and a large-capacity low-performance memory. Data placement on such systems plays a critical role in performance optimization. Existing efforts have explored coarse-grained data placement in applications with dense data structures; however, a thorough study of applications that are based on graph data structures is still missing. This work proposes ATMem—a runtime framework for adaptive granularity data placement optimization in graph applications. ATMem consists of a lightweight profiler, an analyzer using a novel m-ary tree-based strategy to identify sampled and estimated critical data chunks, and a high-bandwidth migration mechanism using a multi-stage multi-threaded approach. ATMem is evaluated in five applications on two HMS hardware, including the Intel Optane byte-addressable NVM and MCDRAM. Experimental results show that ATMem selects 5%-18% data to be placed on high-performance memory and achieves an average of 1.7×-3.4× speedup on NVM-DRAM and 1.2×-2.0× speedup on MCDRAM-DRAM, over the baseline that places all data on the large-capacity memory. On NVM-DRAM, ATMem achieves performance comparable to a full-DRAM system with as low as 9%-54% slowdown.