A Write-Optimized Re-computation Scheme for Non-Volatile Memory

The rise of non-volatile memory (NVM) technology brings many opportunities and challenges to computer storage systems. Compared with DRAM, NVM has the advantages of non-volatility, low energy consumption and high storage density as persistent memory. However, it has the disadvantages of limited erase/write times and high write latency. Therefore, it is necessary to reduce the write operations to the non-volatile main memory to improve system performance and extend the lifetime of NVM. To address this problem, this paper proposes ROD, which is a re-computation scheme based on the out degree of computing nodes. Due to the performance gap between CPU and memory leading to computing resources waste, ROD selectively discards the computed results which should be stored into the NVM and re-computes them when needed. By swapping the storage with computation, ROD reduces the number of writes to the NVM. We have implemented ROD and evaluated it on Gem5 simulator with NVMain. We also implemented the greedy re-computation scheme and the store-only scheme as state-of-the-art work to be compared with ROD. Experimental results from powerstone benchmark show that ROD reduces write operations by 44.3% on average (up to 68.5%) compared with the store-only scheme. The execution time is reduced by 28.1% on average (up to 68.6%) compared with the store-only scheme, and 9.3% on average (up to 19.4%) compared with the greedy scheme.