JCSE, vol. 17, no. 2, pp.80-92, 2023

DOI: http://dx.doi.org/10.5626/JCSE.2023.17.2.80

Accurate Calibration and Scalable Bandwidth Sharing of Multi-Queue SSDs

Hyeongseok Kang and Kanghee Kim
Department of Information Communication, Soongsil University, Seoul, Korea School of Artificial Intelligence Convergence, Soongsil University, Seoul, Korea

Abstract: The emerging multi-queue solid state drives (SSDs) impose two challenges on I/O scheduling in the host operating system. First, the I/O scheduler should give a scalable performance in the number of processor cores to exploit the massive parallelism within the SSD. Second, it should provide performance isolation between the cores so that each core can schedule application I/O streams with a reserved bandwidth share. To cope with these challenges, we propose a novel I/ O scheduler called mqFlashFQ. In mqFlashFQ, for every core to make a scheduling decision in parallel, we use a randomization technique to decentralize the existing FlashFQ algorithm, consequently, significantly reducing the inter-core synchronization overheads. Moreover, to provide a fair bandwidth share on a per-core basis, we present an accurate calibration method that determines the cost of each I/O request in terms of its direction and size. This method is distinguished in that it enables to provide a minimum bandwidth guarantee to each core with no garbage collection. Through our experiments with non-volatile memory express (NVMe) SSD products, we demonstrate that the proposed mqFlashFQ and calibration method give a scalable performance and a fair share of the bandwidth to each core for various I/O workloads.

Keyword: Multi-queue SSDs; I/O scheduling; Fair queuing; Randomization; Request cost calibration

Full Paper:   65 Downloads, 440 View