On Tue, Oct 03, 2017 at 02:21:43AM -0700, Christoph Hellwig wrote: > This looks generally good to me, but I really worry about the impact > on very high iops devices. Did you try this e.g. for random reads > from unallocated blocks on an enterprise NVMe SSD? Looks no such impact, please see the following data in the fio test(libaio, direct, bs=4k, 64jobs, randread, none scheduler) [root@storageqe-62 results]# ../parse_fio 4.14.0-rc2.no_blk_mq_perf+-nvme-64jobs-mq-none.log 4.14.0-rc2.BLK_MQ_PERF_V5+-nvme-64jobs-mq-none.log --------------------------------------- IOPS(K) | NONE | NONE --------------------------------------- randread | 650.98 | 653.15 --------------------------------------- OR: If you worry about this impact, can we simply disable merge on NVMe for none scheduler? It is basically impossible to merge NVMe's request/bio when none is used, but it can be doable in case of kyber scheduler. -- Ming
