Scalio - Scaling up DPU-based JBOF Key-value Store with NVMe-oF Target Offload

Learn about Scalio, a scalable disaggregated key-value store designed to overcome the limitations of current Data-Processing-Unit-based (DPU-based) Just a Bunch of Flash (JBOF) systems in this 15-minute conference presentation from OSDI '25. Discover how researchers from Tsinghua University address the scalability challenges that existing JBOF solutions face when handling increasing numbers of attached SSDs, which traditionally rely heavily on the DPU's CPU for SSD I/O operations. Explore Scalio's innovative approach that offloads SSD I/O operations to the DPU's network I/O capabilities, utilizing both traditional RDMA verbs and the recent hardware optimization of NVMe over Fabrics Target Offload. Understand the system's two-layer design featuring compact in-memory data structures that efficiently handle hot read traffic and manage bursty writes. Examine the critical challenge of maintaining consistency between DRAM states in the DPU and SSD states, and learn about Scalio's RDMA-based cache consistency protocol that ensures linearizability across the disaggregated architecture without relying on hardware cache coherence protocols. Review experimental results demonstrating Scalio's significant improvements in both scalability and throughput, achieving up to 3.3× higher performance compared to existing systems, particularly in high-density SSD configurations, making it a compelling solution for data-intensive applications requiring energy-efficient and cost-effective storage architectures.