Fair I/O scheduler for alleviating read/write interference by forced unit access in flash memory

Abstract

For the past few years, the enterprise Solid State Drives that employ NVM Express are widely used due to their high performance. It is common for multiple tenants and processes to share a single SSD. Providing fair SSD performance for multiple applications has become an important issue. We observe that the write request with FUA flag delays the processing of read request due to SSD internal read/write interference. To alleviate the performance degradation of the read requests, we propose TABS, per-Type fAir Bandwidth I/O Scheduler for NVMe SSDs. TABS determines the fair bandwidth proportional to the maximum bandwidth and I/O issue rate for each type of request. (i) Two-phase Dynamic Scheduling sets fairness goals dynamically according to the I/O patterns of workload. Then, it throttles the FUA write requests to meet the pre-measured fairness goal. (ii) Software-based feedback makes more accurate scheduling possible. By using these techniques, TABS can guarantee fairness between the read and FUA flagged write requests. Finally, compared with the fairness goals, TABS achieves 76% fairness on average and up to 99.5% fairness, while the noop scheduler, the default Linux scheduler, shows 18% fairness.