On 03/28/2017 01:58 PM, Stephen Bates wrote: >>> >>> Thanks for the review Sagi. I’d be OK going with <=0 as the exact >>> match would normally be for minimal IO sizes (where <= and = are the >>> same thing). I will see what other feedback I get and aim to do a >>> respin soon… >> >> No tunables for this, please. There's absolutely no reason why we should >> need it. > > Jens – by this you mean you want to only bucket IO that are exactly > the minimum block size supported by the underlying block device? I was > envisioning we might want to relax that in certain cases (e.g. bucket > 4KB and below going to a 512B device). Sorry, the above was a bit terse. I think a much better solution would be to create a number of buckets (per data direction) and do stats on all of them. The buckets would cover a reasonable range of request sizes. Then when you poll for a given request, we can base our timed number on the data direction AND size of it. You can get pretty far with a few buckets: 512b 4k 8k 16k 32k 64k 128k and you could even have your time estimation function turn these into something sane. Or just use a composite of buckets, if you wish. -- Jens Axboe
