DPDA: A Dynamic Partition-Driven Data Allocation Scheme for Boosting Hybrid SSD Performance
Abstract
QLC NAND flash provides superior storage density and cost-effectiveness at the expense of substantial programming latencies. To improve latency performance, hybrid SSDs incorporate high-speed SLC cache to mitigate the latency overhead of QLC flash. However, the hybrid architecture induces a critical performance bottleneck termed Hot-Write Blocking (HWB): prolonged QLC write operations monopolize flash chips, thereby stalling subsequent hot-data writes to SLC cache. This interference significantly exacerbates queuing latency and severely degrades SLC write performance. To address this issue, we propose a dynamic partition-driven data allocation scheme called DPDA that mitigates the hot-write blocking problem in hybrid architecture SSDs. DPDA partitions the parallel planes of an SSD into hot and cold partitions, and segregates hot and cold requests to the planes of distinct partitions, thereby isolating mutually interfering write streams. Specifically, the scheme implements a data dispatch module to allocate latency-sensitive hot requests to the SLC regions of planes in the hot partition, thereby shielding them from the interference of cold QLC writes. Moreover, to prevent insufficient SLC space in hot partitions, DPDA implements a partition control unit to dynamically swap the planes of hot and cold partitions by space utilization, and then uses the SLC regions of the swapped plane to serve hot requests, thus fully utilizing the SLC resources in hybrid SSDs. We compare DPDA with three state-of-the-art schemes SPA, AMPHIBIAN and RRTS in terms of write waiting time, write latency, and read latency under eight real-world workloads. The experimental results show that compared with the existing schemes, DPDA slashes the hot write waiting time, hot write latency, and average write latency by 52.99%, 51.86%, and 47.55%, respectively.