FAST I/O: QoS Supports for Urgent I/Os in NVMe SSDs.

NVMe SSDs employ lots of I/O submission queues to dispatch I/O requests simultaneously and exploit the parallelism of SSDs actively. Although this architecture can improve the I/O bandwidth and IOPS of NVMe SSDs, the QoS of each I/O request is never guaranteed. Especially, when many I/O requests are pending in the submission queues, I/O requests to be processed urgently can be delayed. In this paper, we present a scheme that handles the urgent I/O requests without a delay. Since the pending I/O requests in the submission queues cannot be controlled in the host, we use the host memory buffer (HMB) as a fast track for processing urgent I/O requests. Instead of sending urgent I/O requests into the SSDs through a legacy I/O path, the proposed scheme removes the latency by directly inserting them in the HMB that can be accessed from both host and SSDs. Through the experiments on our emulator, we show that the proposed scheme could reduce the average and tail latencies significantly.