An Empirical Study on NVM-based Block I/O Caches

Abstract

This paper presents an empirical study on block I/O caches when combining the performance benefits of emerging NVM storages and the cost-effectiveness of secondary storages. Current state-of-the-art I/O caching solutions are designed around the performance characteristics of SSDs, and our study reveals that using them with NVM storages does not fully reap the benefits of I/O devices with near-DRAM latencies. With fast NVM storages, locating data must be handled efficiently, but the sophisticated yet complex data structures used in existing designs impose significant overheads by substantially increasing the hit time. As this design approach is suboptimal for accessing fast I/O devices, we suggest several architectural designs to exploit the performance of NVM storages.