Boosting Burst I/O Performance under Heterogeneous Hybrid Storage Workloads via DUal I/O IssuE Thresholds

I/O schedulers are the key component to maintain high quality of services under homogeneous hybrid storage workloads, e.g., multiple concurrently running sustained workloads. However, the long hardware queue for I/O requests under I/O schedulers, which is utilized to exploit the substantial parallelism of underlying storage devices, induces unbearable I/O latency for the latency-sensitive burst workloads, particularly under heterogeneous hybrid storage workloads, i.e., concurrently running burst and sustained workloads. Based on our dedicated observation that limited slots in the hardware queue are sufficient to maintain a high throughput for the throughput-sensitive sustained workloads consisting of oversized I/O requests, we propose a novel I/O scheduling framework DUET that deploys DUal I/O issuE Thresholds to boost the I/O performance of burst workloads under heterogeneous hybrid workloads. Specifically, DUET utilizes a high threshold to limit the total amount of data requested by the sustained workloads in the hardware queue when only sustained workloads are actively running. According to our observation, a proper high threshold not only ensures a high throughput for sustained workloads, but also significantly reduces the queuing time for the I/O requests of upcoming burst workloads. To further reduce the I/O latency of burst workloads, DUET adopts a low threshold when sustained and burst workloads are simultaneously present. Through adjusting this low threshold, we can conveniently trade off throughput of sustained workloads against I/O latency of burst workloads. Results from prototyping experiments demonstrate the effectiveness and efficiency of DUET.