Re: stochastic fair queueing in the elevator [Re: [BENCHMARK] 2.4.20-ck3

[Nick Piggin]

Andrew Morton wrote:

>Nick Piggin <[e-mail]> wrote:
>
>>That is what I can't understand. Movement of the disk head should
>>be exactly the same in either situation and 128K is not exactly
>>a pitiful request size - so it suggests a quirk somewhere. It
>>is not as if the disk has to be particularly smart or know a
>>lot about the data in order to optimise the head movement for
>>a load like this.
>>
>
>Yes, that's a bit odd.  Some reduction in CPU cost and bus
>traffic, etc would be expected.   Could be that sending out a
>request which is larger than a track is saving a rev of the disk
>for some reason.
>
Shouldn't be. Even at 128KB readahead we should always have
outstanding requests against the disk in a streaming read
scenario, right? Maybe if the track buffers are bigger than
128K?

Is there a magic number above which you see the improvement,
Andrea? Or does it steadily climb?