Re: [PATCH 1/2]block cfq: make queue preempt work for queues from different workload

[Jeff Moyer <jmoyer@redhat.com>]

Jens Axboe <jaxboe@fusionio.com> writes:

> On 2011-01-14 05:44, Shaohua Li wrote:
>> 2011/1/12 Shaohua Li <shaohua.li@intel.com>:
>>> Hi,
>>> On Wed, Jan 12, 2011 at 05:07:47AM +0800, Corrado Zoccolo wrote:
>>>> Hi Shaohua,
>>>> On Tue, Jan 11, 2011 at 9:51 AM, Shaohua Li <shaohua.li@intel.com> wrote:
>>>>> I got this:
>>>>>             fio-874   [007]  2157.724514:   8,32   m   N cfq874 preempt
>>>>>             fio-874   [007]  2157.724519:   8,32   m   N cfq830 slice expired t=1
>>>>>             fio-874   [007]  2157.724520:   8,32   m   N cfq830 sl_used=1 disp=0 charge=1 iops=0 sect=0
>>>>>             fio-874   [007]  2157.724521:   8,32   m   N cfq830 set_active wl_prio:0 wl_type:0
>>>>>             fio-874   [007]  2157.724522:   8,32   m   N cfq830 Not idling. st->count:1
>>>>> cfq830 is an async queue, and preempted by a sync queue cfq874. But since we
>>>>> have cfqg->saved_workload_slice mechanism, the preempt is a nop.
>>>>> Looks currently our preempt is totally broken if the two queues are not from
>>>>> the same workload type.
>>>>> Below patch fixes it. This will might make async queue starvation, but it's
>>>>> what our old code does before cgroup is added.
>>>> have you measured latency improvements by un-breaking preemption?
>>>> AFAIK, preemption behaviour changed since 2.6.33, before cgroups were
>>>> added, and the latency before the changes that weakened preemption in
>>>> 2.6.33 was far worse.
>>> Yes. I'm testing a SD card for MeeGo. The random write is very slow (~12k/s) but
>>> random read is relatively fast > 1M/s.
>>>
>>> Without patch:
>>> write: (groupid=0, jobs=1): err= 0: pid=3876
>>>  write: io=966656 B, bw=8054 B/s, iops=1 , runt=120008msec
>>>    clat (usec): min=5 , max=1716.3K, avg=88637.38, stdev=207100.44
>>>     lat (usec): min=5 , max=1716.3K, avg=88637.69, stdev=207100.41
>>>    bw (KB/s) : min=    0, max=   52, per=168.17%, avg=11.77, stdev= 8.85
>>> read: (groupid=0, jobs=1): err= 0: pid=3877
>>>  read : io=52516KB, bw=448084 B/s, iops=109 , runt=120014msec
>>>    slat (usec): min=7 , max=1918.5K, avg=519.78, stdev=25777.85
>>>    clat (msec): min=1 , max=2728 , avg=71.17, stdev=216.92
>>>     lat (msec): min=1 , max=2756 , avg=71.69, stdev=219.52
>>>    bw (KB/s) : min=    1, max= 1413, per=66.42%, avg=567.22, stdev=461.50
>>>
>>> With patch:
>>> write: (groupid=0, jobs=1): err= 0: pid=4884
>>>  write: io=81920 B, bw=677 B/s, iops=0 , runt=120983msec
>>>    clat (usec): min=13 , max=742976 , avg=155694.10, stdev=244610.02
>>>     lat (usec): min=13 , max=742976 , avg=155694.50, stdev=244609.89
>>>    bw (KB/s) : min=    0, max=   31, per=inf%, avg= 8.40, stdev=12.78
>>> read: (groupid=0, jobs=1): err= 0: pid=4885
>>>  read : io=133008KB, bw=1108.3KB/s, iops=277 , runt=120022msec
>>>    slat (usec): min=8 , max=1159.1K, avg=164.24, stdev=9116.65
>>>    clat (msec): min=1 , max=1988 , avg=28.34, stdev=55.81
>>>     lat (msec): min=1 , max=1989 , avg=28.51, stdev=57.51
>>>    bw (KB/s) : min=    2, max= 1808, per=51.10%, avg=1133.42, stdev=275.59
>>>
>>> Both read latency/throughput has big difference with the patch, but write
>>> gets starvation.
>> Hi Jens and others,
>> How do you think about the patch?
>
> I think the patch is good. If preemption in some cases makes things
> worse, then we need to look into those. That's a separate issue.

I think things are getting pretty messy.  Who would have guessed that
the saved_workload_slice would affect preemption?  If a queue is to
preempt another, can't we make that a bit more explicit?  I'm still
trying to walk through the code to figure out how this ends up
happening, as the patch description leaves a lot out of the picture.

Cheers,
Jeff