MLCache: a space-efficient cache scheme based on reuse distance and machine learning for NVMe SSDs

Non-volatile memory express (NVMe) solid-state drives (SSDs) have been widely adopted in emerging storage systems, which can provide multiple I/O queues and high-speed bus to maximize high data transfer rate. NVMe SSD use streams (also called "Multi-Queue") to store related data in associated locations or for other performance enhancements. The on-board DRAM cache inside NVMe SSDs can efficiently reduce the disk accesses and extend the lifetime of SSDs, thus improving the overall efficiency of the storage systems. However, in previous studies, such SSD cache has been only used as a shared cache for all streams or a statically partitioned cache for each stream, which may seriously degrade the performance-per-stream and underutilize the valuable cache resources.

In this paper, we present MLCache, a space-efficient shared cache management scheme for NVMe SSDs, which maximizes the write hit ratios, as well as enhances the SSD lifetime. We formulate cache space allocation as a machine learning problem. By learning the impact of reuse distance on cache allocation, we build a workload specific neural network model. At runtime, MLCache continuously monitors the reuse distance distribution for the neural network module to obtain space-efficient allocation decisions. Experimental results show MLCache improves the write hit ratio of the SSD by 24% compared to baseline, and achieves response time reduction by 13.36% when compared with baseline. MLCache is 96% similar to the ideal model.