PRIOR-I/O: Real-time Performance Improvement through Dynamic Resource Allocation for Virtual I/O

With the widespread adoption of virtualization, ensuring predictable I/O latency in consolidated virtualized environments has become crucial. Predictable virtual I/O is essential for enhancing system responsiveness and maintaining target latency satisfaction when multiple VMs on the same host compete for shared storage bandwidth. However, the complex architecture of virtualized systems introduces significant challenges, as I/O operations compete not only with tasks within the same virtual machine (VM) but also across multiple VMs sharing the same host resources. This paper introduces PRIOR-I/O, a novel framework designed to deliver predictable real-time I/O performance in virtualized environments. A system-level monitor is implemented to dynamically track I/O workloads and task requirements, enabling real-time adjustments. By utilizing either an analytical model or a learning-based neural network model, PRIOR-I/O accurately predicts tail I/O latency under varying conditions. These predictions drive a dynamic priority assignment mechanism that optimizes resource allocation based on task criticality and workload demands. The effectiveness of the framework is demonstrated through extensive experiments, showing a 41.25%–47.29% improvement in real-time performance , particularly in dynamic I/O scenarios where workloads fluctuate unpredictably. Notably, PRIOR-I/O achieves these gains without kernel modifications or specialized hardware, ensuring broad compatibility across diverse virtualization platforms. The results highlight its applicability in consolidated multi-VM deployments, where maintaining predictable I/O performance is paramount.