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

[Nick Piggin]

Andrea Arcangeli wrote:

>On Mon, Feb 10, 2003 at 10:24:57PM +1100, Nick Piggin wrote:
>
>>Andrea Arcangeli wrote:
>>
>>
>>>On Mon, Feb 10, 2003 at 09:40:34PM +1100, Nick Piggin wrote:
>>>
>>>
>>>>I don't know too much about SCSI stuff, but if driver / wire / device
>>>>overheads were that much higher at 128K compared to 512K I would
>>>>think something is broken or maybe optimised badly.
>>>>
>>>>
>>>I guess it's also a matter of the way the harddisk can serve the I/O if
>>>it sees it all at the same time, not only the cpu/bus protocol after all
>>>minor overhead.  Most certainly it's not a software mistake in linux
>>>that the big commands runs that much faster. Again go check the numbers
>>>in bigbox.html between my tree, 2.4 and 2.5 in bonnie read sequential,
>>>to see the difference between 128k commands and 512k commands with
>>>reads, these are facts.  (and no writes and no seeks here)
>>>
>>>
>>Yes it is very clear from the numbers that your tree is more than
>>150% the speed for reads. As I said I don't know too much about
>>
>
>correct, that's the huge improvement I get in the read sequential case
>(i.e. bonnie), which is a crucial common workload.
>
Yep

>
>
>>SCSI, but it is very interesting that writes don't get a noticable
>>improvement although they would be using the bigger request sizes
>>too, right? Something is causing this but the cpu, bus, wire
>>
>
>It's the readahead in my tree that allows the reads to use the max scsi
>command size. It has nothing to do with the max scsi command size
>itself.
>
>writes don't need readahead to use the max command size, they always
>used it since the first place, so they can't go even faster, they never
>had a problem.
>
Yes I am an idiot, I don't know why I said that :P

>
>
>It's by fixing readahead that reads gets boosted. this has nothing to do
>with writes or the max_sectors itself.
>
>You can wait 10 minutes and still such command can't grow.  This is why
>claiming anticipatory scheduling can decrease the need for readahead
>doesn't make much sense to me, there are important things you just can't
>achieve by only waiting.
>
Look at the few simple tests I have been posting - it clearly
indicates the need is decreased. I did say however that it would
be subject to CPU/bus/device overheads. I did not realiase
SCSI setups would behave like this. From a purely disk head
perspective it does nullify the need for readahead (though
it is obivously still needed for other reasons).