On Wed, May 6, 2015 at 11:14 AM, Ming Lei <ming.lei@xxxxxxxxxxxxx> wrote: > On Wed, May 6, 2015 at 12:55 AM, Tejun Heo <tj@xxxxxxxxxx> wrote: >> Hello, Ming. >> >> On Tue, May 05, 2015 at 10:46:10PM +0800, Ming Lei wrote: >>> On Tue, May 5, 2015 at 9:59 PM, Tejun Heo <tj@xxxxxxxxxx> wrote: >>> > It's a bit weird to hard code this to 16 as this effectively becomes a >>> > hidden bottleneck for concurrency. For cases where 16 isn't a good >>> > value, hunting down what's going on can be painful as it's not visible >>> > anywhere. I still think the right knob to control concurrency is >>> > nr_requests for the loop device. You said that for linear IOs, it's >>> > better to have higher nr_requests than concurrency but can you >>> > elaborate why? >>> >>> I mean, in case of sequential IO, the IO may hit page cache a bit easier, >>> so handling the IO may be quite quick, then it is often more efficient to >>> handle them in one same context(such as, handle one by one from IO >>> queue) than from different contexts(scheduled from different worker >>> threads). And that can be made by setting a bigger nr_requests(queue_depth). >> >> Ah, so, it's about the queueing latency. Blocking the issuer from >> get_request side for the same level of concurrency would incur a lot >> longer latency before the next IO can be dispatched. The arbitrary 16 >> is still bothering but for now it's fine I guess, but we need to >> revisit the whole thing including WQ_HIGHPRI thing. Maybe it made >> sense when we had only one thread servicing all IOs but w/ high >> concurrency I don't think it's a good idea. > > Yes, I was thinking about it too, but concurrency can improve > random I/O throughput a lot in my tests. Thinking of it further, the problem becomes very similar with 'Non-blocking buffered file read operations'[1] which was discussed last year. If the read IO can be predicted as buffered I/O, we handle it in single thread, otherwise handle it concurrently, and this approach should be more efficient if possible, I think. But I still prefer to dio/aio approach because double cache can be avoided, which is a big win in my previous tests. [1], https://lwn.net/Articles/612483/ Thanks, Ming Lei -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html