Online Application Guidance for Heterogeneous Memory Systems

As scaling of conventional memory devices has stalled, many high-end computing systems have begun to incorporate alternative technologies meet performance goals. Since these present distinct advantages and tradeoffs compared DDR* SDRAM, such as higher bandwidth with lower capacity or vice versa, they are typically packaged alongside SDRAM in a heterogeneous architecture. To utilize the different types efficiently, new data management strategies needed match application usage best available technology. However, current proposals for managing memories limited, because either (1) do not consider high-level behavior when assigning (2) require separate program execution (with representative input) collect information about how uses resources. This work presents toolset address limitations existing approaches complex memories. It extends runtime layer automated monitoring routines that assign tier based on previous usage, without any need source code modification profiling run. evaluates this approach state-of-the-art server platform both DDR4 non-volatile Intel Optane DC memory, using memory-intensive high-performance (HPC) applications well standard benchmarks. Overall, results show improves significantly unguided across variety workloads system configurations. The HPC exhibit largest benefits, speedups ranging from 1.4× 7× cases. Additionally, we achieves similar comparable offline profiling-based after short startup period, requiring analysis steps.