On 2023-02-06 21:20, Christoph Hellwig wrote: > On Fri, Feb 03, 2023 at 04:00:06PM +0530, Pankaj Raghav wrote: >> move to blk-mq based request processing as brd is one of the few drivers >> that still uses submit_bio interface. The changes are pretty trivial >> to start using blk-mq. The performance increases up to 125% for direct IO >> read workloads. There is a slight dip in performance for direct IO write >> workload but considering the general performance gain with blk-mq >> support, it is not a lot. > > Can you find out why writes regress, and what improves for reads? > Definitely. I tried to do similar experiments in null blk with submit_bio & blk-mq backend and noticed a similar pattern in performance. I will look into it next week as I am OOO rest of the week. > In general blk-mq is doing a lot more work for better batching, but much > of that batching should not matter for a simple ramdisk. So the results > look a little odd to me, and extra numbers and explanations would > really help. Let me know if you think I am missing something in the code that can cause this behavior! Thanks.