QoS, Fairness and Isolation in One Formula

Abstract

Performance isolation remains challenging in multi-tenant cloud storage. In private clouds like VMware vSAN, “noisy neighbor” effects typically occur under bursty traffic.

We propose a lightweight, non-intrusive resource allocation formula that eliminates reliance on single bottleneck controllers or specialized device features. This design schedules per-VM I/Os using hypervisor throttling, preserving FCFS service assumption at device ends. A unified linear programming formula runs in a feedback loop to solve optimal IOPS rate allocation per virtual disk, maximizing throughput under QoS and device capacity constraints.

Simulations show effective I/O throttling, burst traffic isolation at the source, reduced destination congestion, and significantly lower end-to-end latency.

Available at GitHub.