Improving I/O Performance of Large-Page Flash Storage Systems Using Subpage-Parallel Reads

Abstract

Handling small read requests is important on large-page flash storage systems because small reads tend to waste the read bandwidth. We present a system level solution that avoids wasting the read bandwidth based on a new page read operation, called the subpage-parallel read (SPREAD). SPREAD achieves an optimal latency for a small read request by reading requested subpages only. By improving the read performance of applications as well as garbage collection, SPREAD improves the overall I/O performance. Experimental results show that an SPREAD-aware FTL can improve the IOPS and read latency by up to 122% and 56%, respectively.