On 25/01/2021 17:04, Jens Axboe wrote: > On 1/25/21 9:56 AM, Pavel Begunkov wrote: >> On 25/01/2021 16:08, Jens Axboe wrote: >>> On 1/25/21 4:42 AM, Pavel Begunkov wrote: >>> >>> Applied 1-2 for now. Maybe the context state is workable for the >>> state, if the numbers are looking this good. I'll try and run >>> it here too and see what I find. >> >> I believe I've seen it jumping from ~9.5M to 14M with turbo boost >> and not so strict environment, but this claim is not very reliable. > > Would be nice to see some numbers using eg a network echo bench, > or storage that actually does IO. Even null_blk is better than just > a nop benchmark, even if that one is nice to use as well. But better > as a complement than the main thing. I think direct I/O null_blk won't cut because it completes through a callback, but maybe some buffered rw. Yeah, good idea to try something more realistic. -- Pavel Begunkov
