MQSim 2.0: A Framework for Realistic Studies of AI-Era SSDs and Disaggregated Storage

Abstract

Modern storage systems are being re-architected across several coupled layers: dense QLC media with pSLC caching and one-shot programming, host-directed placement (FDP , ZNS), disaggregated NVMe-oF data paths, and flash-management logic that can span SSDs, hosts, and DPUs, a regime driven by AI serving and dense multi-tenant cloud storage alike. Existing SSD simulators model storage through device-centric abstractions or support these mechanisms along separate axes, making their end-to-end interactions difficult to study in one framework. We present MQSim 2.0 , a composable storage-system simulator that supports direct-attached and disaggregated deployments and uses SplitFTL to model flash-management decomposition across the device and an upstream controller. We validate MQSim 2.0 against a TLC NVMe Gen5 SSD and a real 24-SSD DPU-based storage enclosure, achieving throughput/latency MAPE of 7.9%/3.5% and 4.4%/4.8%, respectively. We then use MQSim 2.0 to study a disaggregated KV-cache design combining retention-aware placement, ZNS backends, and DPU-resident control, which reduces write amplification by 48% and P99 latency by 31% over the evaluated JBOF baseline.