Rectifying pitfalls in the performance evaluation of flash solid-state drives

Abstract

Today's data storage systems are increasingly adopting flash-based solid-state drives (SSD), in which new data is written out-of-place. The space occupied by invalidated data is reclaimed by means of a garbage-collection process. This involves additional write operations that result in write amplification, which negatively affects the performance, endurance, and lifetime of SSDs. Several garbage-collection schemes have been proposed in the literature and corresponding models have been developed for assessing their efficiency. We demonstrate that some of these publications arrive at conflicting results. We establish that the discrepancies identified are due to pitfalls in the modeling and analysis of some of the basic garbage-collection schemes. We effectively resolve these discrepancies by rectifying the pitfalls and developing proper analytical models that yield accurate results. We obtain new results for the circular scheme that are subsequently used to develop a new accurate model for the windowed greedy garbage-collection scheme. Results of theoretical and practical importance are analytically derived and experimentally confirmed. They demonstrate that, as the window size decreases, the write amplification increases, illustrating the tradeoff between computation time and write amplification. The results also show that, under certain conditions, the write amplification of the simple circular scheme can be similar to that of the optimum, albeit more complex greedy garbage-collection scheme. In this case, the write amplification for the windowed greedy scheme is essentially found to be independent of the window size. © 2014 Elsevier B.V. All rights reserved.