Contention-free fair queuing for high-speed storage with RAID-0 architecture

Controlling the latency and throughput is a crucial part of designing an I/O scheduler that can provide guaranteed quality-of-service of a shared storage resource to multiple competing applications. Previously proposed I/O schedulers have mainly focused on ensuring proportional bandwidth allocation and a user-defined latency. However, the algorithms are not optimized for high-speed storage with restricted resources such as solid-state drives (SSD), because the storage performance can widely vary depending on the I/O patterns or the delays associated with different I/O operations as well as the I/O size. To solve these problems, we propose contention-free fair queuing (CFFQ) that accurately predicts the processing time of I/O requests considering the internal constraints of the storage, such as static striping and limited concurrency. Moreover, CFFQ compensates for the loss time of unselected I/O requests to ensure the bounded worst-case latency. In experiments, we have evaluated the guaranteed QoS of CFFQ versus the previous I/O scheduler (SFQ(D)) for diverse workloads. The experiments show that the throughput of CFFQ versus SFQ(D) is increased by up to 65% according to the workload while guaranteeing fairness and the worst-case latency.