An Evaluation Methodology for Exascale Memory Systems

Evaluating potential Exascale memory systems is difficult because of the scale of the target system. While full-workload simulation techniques work at the core and cache level granularity, the shielding effects of the cache require extremely large simulation times to simulate a significant number of memory accesses to assess performance. The current approach to simulate such systems is the use of memory trace files, but these are difficult to generate and are inflexible in their ability to capture important data dependencies, which may be important in assessing revolutionary memory subsystems. Our methodology takes a two-pronged approach to estimate power, performance, and resiliency of large scale memory systems. We first improve the trace-based approach to measure power and resiliency, and then take a hardware approximation approach to evaluate performance for proxy-applications. Our design is useful for exploring traditional DRAM, 3D-Stacked DRAM, non-volatile memories, or any combination