TITLE: Towards Service-Level Objectives: A Novel Fairness Measurement of SR-IOV SSDs for Large-scale Computing Systems.

ABSTRACT: The increasing deployment of large-scale AI training and inference workloads has tightened service-level latency objectives and amplified the role of storage in end-to-end responsiveness. In this landscape, NVMe-based Solid-State Drives (SSDs) are a critical bottleneck, especially when shared across multiple tenants in virtualized cloud environments. While Single Root I/O Virtualization (SR-IOV) has emerged as a hardware-level solution to reduce virtualization overhead and isolate per-tenant I/O paths, current evaluation methodologies fail to quantify fairness in tail-latency distribution among virtual machines (VMs). This paper addresses this gap by introducing a device-level fairness metric tailored for SR...