LEPA: Latency estimation-based page allocation for improving SSD performance

Abstract

Modern solid state drives (SSDs) based on NAND Flash Memory (NFM) have multi-level parallel resources to enhance their I/O performance. The page allocation policy, which is responsible for allocating logical pages to physical parallel resources, directly affects the efficiency of SSD parallelism utilization. Traditionally, the load-balancing page allocation policy relies on the number of commands rather than their actual latency. However, this policy fails to effectively balance the execution latency skew among the various parallel units of SSD, leading to parallelism loss.

To address these problems, we propose a load-balancing method based on latency estimation called LEPA. Instead of relying on the number of commands, LEPA estimates the waiting latency of commands in pending to determine the load. Our experimental results indicate that our latency estimation-based load balancing page allocation significantly reduces the die load skew (by 85.9% on average) caused by the inaccurate estimations of traditional methods, with minimal overhead. Moreover, LEPA demonstrates an 8.0% improvement in plane-level parallelism. As a result, the proposed method effectively improves SSD system I/O performance and reduces request response time by 14.7%.