vPFS: Bandwidth Virtualization of Parallel Storage Systems

This paper presents vPFS, a new parallel file system performance management approach to support the allocation of shared storage bandwidth on a per-application basis. Existing parallel file systems are unable to differentiate I/O requests from different applications and meet per-application bandwidth requirements. This limitation presents an increasing hurdle for applications to achieve their desired Quality of Service (QoS) as high-performance computing (HPC) systems continue to scale up. This paper proposes vPFS, a new solution to address this challenge through a bandwidth virtualization layer for parallel file systems. vPFS employs user-level parallel file system proxies to interpose requests between native parallel file system clients and servers to schedule parallel I/Os from different applications according to configurable bandwidth management policies. The vPFS approach is applicable to different parallel file storage systems and can integrate transparently with existing as well as future HPC systems. The utility and performance of vPFS is demonstrated with a prototype implementation which virtualizes PVFS2, a widely used parallel file system. An enhanced SFQ-based scheduler is enabled for proportional sharing of parallel storage based on its unique characteristics (parallel stripped I/Os) and requirement (high throughput). The enhancements include new thresholdand layout-driven scheduling synchronization schemes, which reduce the global communication overhead while delivering totalservice proportional sharing. Experimental evaluation using typical HPC benchmarks (IOR, NPB BTIO) shows that the overhead due to virtualization in vPFS is small (less than 3% slowdown in I/O throughput). It also shows that vPFS can achieve good proportional bandwidth sharing for competing applications with diverse I/O patterns (> 97% of the target sharing ratio).