LPNS - Scalable and Latency-Predictable Local Storage Virtualization for Unpredictable NVMe SSDs in Clouds

Explore a 23-minute conference talk from USENIX ATC '23 that introduces LPNS, a novel local storage virtualization system designed to provide scalable and latency-predictable performance for unpredictable NVMe SSDs in cloud environments. Delve into the challenges of achieving latency predictability in public clouds and learn how LPNS addresses device-level interference between multi-tenant virtualized devices. Discover the innovative features of LPNS, including reliable self-feedback control, flexible I/O queue and command scheduling, and scalable polling design. Examine the deterministic network calculus-based formalization method used to establish upper bounds for virtualized device latency. Gain insights into LPNS's performance improvements, including up to 18.72× latency optimization compared to mainstream NVMe virtualization and up to 1.45× additional throughput over state-of-the-art storage latency control systems. Presented by researchers from Shanghai Jiao Tong University, this talk offers valuable insights for cloud storage professionals and researchers interested in advancing storage virtualization techniques.