Performance Evaluation on NVMM Emulator Employing Fine-Grain Delay Injection

The emerging technology of byte-addressable nonvolatile memory chips is expected to enable larger main memory and lower power consumption than the traditional DRAM. It also realizes durable data structure without ordinary file systems. However, while enumerating the advantages of nonvolatile main memory (NVMM), its write-time expensive latency and higher energy consumption in comparison with a DRAM must be considered. These special characteristics of NVMM require new compiler techniques and OS support as well as new memory architectures. Several NVMM emulators built on real machines have been proposed to facilitate those software and hardware researches. Their designs were originally based on a simple coarse-grain delay model that injected additional clock cycles in the read and write requests sent to the memory controller. However, they could not utilize bank-level parallelism and row-buffer access locality, relied on by today's memory modules, to exploit their performance. Therefore, a fine-grain delay model was recently proposed where the delay is injected for the primitive memory operations issued by the memory controller. In this paper, we implement both the coarse-grain and the fine-grain delay models on an SoC-FPGA board along with the use of Linux kernel modifications and several runtime functions. Then, the program behavior differences between two models are evaluated with SPEC CPU programs. The fine-grain model reveals the program execution time is influenced by the frequency of NVMM memory requests rather than the cache hit ratio. Bank-level parallelism and row-buffer access locality also affect the memory access delay, and the fine-grain model shows lower execution time for four of fourteen programs than the coarse-grain even when the former has longer total write latency.