Re: [patch 3/3]cfq-iosched: don't idle if a deep seek queue is slow

[Shaohua Li <shaohua.li@intel.com>]

On Tue, 2010-11-09 at 10:28 +0800, Vivek Goyal wrote:
> On Tue, Nov 09, 2010 at 09:36:42AM +0800, Shaohua Li wrote:
> > On Mon, 2010-11-08 at 23:06 +0800, Jens Axboe wrote:
> > > On 2010-11-08 15:20, Vivek Goyal wrote:
> > > > On Mon, Nov 08, 2010 at 10:07:25AM +0800, Shaohua Li wrote:
> > > >> If a deep seek queue slowly deliver requests but disk is much faster, idle
> > > >> for the queue just wastes disk throughput. If the queue delevers all requests
> > > >> before half its slice is used, the patch disable idle for it.
> > > >> In my test, application delivers 32 requests one time, the disk can accept
> > > >> 128 requests at maxium and disk is fast. without the patch, the throughput
> > > >> is just around 30m/s, while with it, the speed is about 80m/s. The disk is
> > > >> a SSD, but is detected as a rotational disk. I can configure it as SSD, but
> > > >> I thought the deep seek queue logic should be fixed too, for example,
> > > >> considering a fast raid.
> > > >>
> > > > 
> > > > Hi Shaohua,
> > > > 
> > > > So looks like you are trying to cut down queue idling in the case when
> > > > device is fast and idling hurts. That's a noble goal, just that detetction
> > > > of this condition only for deep queues does not seem to cover lots of
> > > > cases. Manually one can set slice_idle=0 to handle this situation.
> > > > 
> > > > What about if you have lots of sequential queues (not deep) and they all
> > > > will still idle.
> > > > 
> > > > Secondly, what if driver is just buffering lots of requests in its device
> > > > queue and not necessarily device is processing the reuqests faster.
> > > 
> > > That is not a valid concern, a driver should never extract more than it
> > > can process (pretty much) immediately.
> > > 
> > > > So I think it is a good idea to cut down on idling if we can find that
> > > > underlying device is fast and idling on queue might hurt more. But
> > > > discovering this only using deep queues does not sound very appleaing to
> > > > me. This is help only a particular workload which is driving deep queues.
> > > > So if there was a generic mechanism to tackle this, that would be much
> > > > better.
> > > 
> > > Agree, we could use better metrics for this.
> > Agree we'd better have a better method to measure device speed, but this
> > seems not easy. Even in a fast device, a request might take long time to
> > finish when NCQ is enabled. Before we have generic mechanism, we still
> > need fix some particular cases.
> 
> Do you have a real workload for this case or it is just one of the synthetic
> workload simulated using fio?
No, no real workload. We do a lot of fio tests with different scripts,
this is from one of our tests.