实验室硕士生夏豪等人联合撰写的论文《HEAL: A Hierarchical Approach to Shortening Block Open Time on Multi-Tenant SSDs》被《 IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems》录用。 IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems为CCF 推荐A类国际期刊。论文将于2026年正式发表。
Abstract—Multi-tenant SSDs commonly employ isolation to provide predictable QoS. However, block-level (per-tenant) isola tion itself prolongs block open time (the interval from when the first page is programmed until the last page is programmed): by forbidding cross-tenant write aggregation within a block, it fragments write streams and lowers each block’s effective fill rate; tenants with low write rates then fill blocks slowly, and modern 3D NAND’s deep parallelism keeps many blocks open concurrently. As blocks remain open longer, threshold-voltage distributions are perturbed and raw bit error rates rise. The controller first relies on ECC to absorb the additional raw errors; if decoding with the current read reference voltages still fails, it invokes read retry with retuned reference voltages, which inflates read latency. Persistently elevated RBER also raises background re fresh pressure, requiring extra migration and rewrite operations to preserve reliability. 3D NAND scaling (larger pages/blocks) further magnifies exposure, and physical load imbalance across planes can strand "cold" planes, prolonging block open time regardless of the tenant policy. We present HEAL, a lightweight, device-side hierarchical approach that shortens block open time. HEAL coordinates cross-plane balancing, tiered hybrid block allocation, and adaptive garbage collection to preserve parallelism and stability. Across 18 real-world workloads, HEAL reduces block open time by 29% at 16 tenants and improves read latency by up to 7.21% and throughput by up to 3.2%.