Improving I/O Performance and Fairness in NVMe SSDs With Pooling Portions of Cache Partitions

NVMe SSDs have become mainstream storage devices in today’s computing systems, due to their high throughput and ultra-low latency. It has been observed that the impact of interference among all concurrently running streams (i.e., I/O workloads) on their overall responsiveness differs significantly in multi-stream SSDs, resulting in unfairness. This paper proposes a cache division management scheme built on top of the evenly partition scheme for NVMe SSDs, to enhance I/O responsiveness without consciously sacrificing fairness. To this end, we first build a mathematical model to directly cut portions from the Local cache partitions allocated to concurrently running streams, considering their run-time performance measures. Then, our approach pools these portions together for the use of all streams. As a result, each stream has its corresponding Local cache space for ensuring fairness, meanwhile the pooled Global cache space is shared by all streams for enhancing I/O responsiveness. Tracedriven simulation experiments demonstrate that our proposal reduces the overall I/O latency by up to 24.4%, and improve the measure of fairness by 2.5× on average, in contrast to existing cache management schemes for NVMe SSDs.