PARDA: Proportional allocation of resources for distributed storage access

Abstract

Rapid adoption of virtualization technologies has led to increased utilization of physical resources, which are multiplexed among numerous workloads with varying demands and importance. Virtualization has also accelerated the deployment of shared storage systems, which offer many advantages in such environments. Effective resource management for shared storage systems is challenging, even in research systems with complete end-to-end control over all system components. Commercially-available storage arrays typically offer only limited, proprietary support for controlling service rates, which is insufficient for isolating workloads sharing the same storage volume or LUN. To address these issues, we introduce PARDA, a novel software system that enforces proportional-share fairness among distributed hosts accessing a storage array, without assuming any support from the array itself. PARDA uses latency measurements to detect overload, and adjusts issue queue lengths to provide fairness, similar to aspects of flow control in FAST TCP. We present the design and implementation of PARDA in the context of VMware ESX Server, a hypervisor-based virtualization system, and show how it can be used to provide differential quality of service for unmodified virtual machines while maintaining high efficiency. We evaluate the effectiveness of our implementation using quantitative experiments, demonstrating that this approach is practical.