4Z-SSD: A Semantic-Aware Four-Zone SSD Architecture for Balancing Write Amplification and Flash-Level Parallelism

Abstract

Modern SSDs face a fundamental trade-off between reducing write amplification and maximizing flash-level parallelism. Conventional FTL schemes improve resource utilization through aggressive data distribution, but incur higher garbage collection overhead by intermixing data with different update behaviors. In contrast, semantic-aware schemes reduce write amplification through workload-driven clustering, but often sacrifice internal parallelism and introduce resource contention. This paper presents 4Z-SSD, a semantic-aware four-zone SSD architecture that jointly optimizes garbage collection efficiency and flash-level parallelism. 4Z-SSD classifies writes based on temperature and access pattern, organizing data into four semantic zones: Hot-Sequential, Hot-Random, Cold-Sequential, and Cold-Random. It further integrates workload-aware buffering and multi-plane-friendly allocation/erase to reduce unnecessary valid-page migration while preserving aligned flash operations. Experimental results show that 4Z-SSD reduces write amplification by up to 11.3% compared with CATDOG and improves throughput by up to 1.95× compared with ReCAFTL.