Re: [PATCH]cfq-iosched: don't take requests with long distence asclose

From: Jens Axboe
Date: Mon Dec 28 2009 - 07:17:07 EST


On Mon, Dec 28 2009, Shaohua Li wrote:
> On Mon, Dec 28, 2009 at 05:11:25PM +0800, Corrado Zoccolo wrote:
> > Hi Shaohua,
> > On Mon, Dec 28, 2009 at 9:46 AM, Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> > > On Mon, Dec 28, 2009 at 04:36:39PM +0800, Corrado Zoccolo wrote:
> > >> Hi Shaohua,
> > >> On Mon, Dec 28, 2009 at 3:03 AM, Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> > >> > On Fri, Dec 25, 2009 at 06:16:27PM +0800, Corrado Zoccolo wrote:
> > >> >> Hi Shaohua,
> > >> >> On Thu, Dec 24, 2009 at 1:55 AM, Shaohua Li <shaohua.li@xxxxxxxxx> wrote:
> > >> >> > df5fe3e8e13883f58dc97489076bbcc150789a21
> > >> >> > b3b6d0408c953524f979468562e7e210d8634150
> > >> >> > The coop merge is too aggressive. For example, if two tasks are reading two
> > >> >> > files where the two files have some adjecent blocks, cfq will immediately
> > >> >> > merge them. cfq_rq_close() also has trouble, sometimes the seek_mean is very
> > >> >> > big. I did a test to make cfq_rq_close() always checks the distence according
> > >> >> > to CIC_SEEK_THR, but still saw a lot of wrong merge. (BTW, why we take a long
> > >> >> > distence far away request as close. Taking them close doesn't improve any thoughtput
> > >> >> > to me. Maybe we should always use CIC_SEEK_THR as close criteria).
> > >> >> Yes, when deciding if two queues are going to be merged, we should use
> > >> >> the constant CIC_SEEK_THR.
> > >> >
> > >> > seek_mean could be very big sometimes, using it as close criteria is meanless
> > >> > as this doen't improve any performance. So if it's big, let's fallback to
> > >> > default value.
> > >>
> > >> meanless -> meaningless (also in the comment within code)
> > > oops
> > >
> > >> > Signed-off-by: Shaohua Li<shaohua.li@xxxxxxxxx>
> > >> >
> > >> > diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> > >> > index e2f8046..8025605 100644
> > >> > --- a/block/cfq-iosched.c
> > >> > +++ b/block/cfq-iosched.c
> > >> > @@ -1682,6 +1682,10 @@ static inline int cfq_rq_close(struct cfq_data *cfqd, struct cfq_queue *cfqq,
> > >> >        if (!sample_valid(cfqq->seek_samples))
> > >> >                sdist = CFQQ_SEEK_THR;
> > >> >
> > >> > +       /* if seek_mean is big, using it as close criteria is meanless */
> > >> > +       if (sdist > CFQQ_SEEK_THR)
> > >> > +               sdist = CFQQ_SEEK_THR;
> > >> > +
> > >> >        return cfq_dist_from_last(cfqd, rq) <= sdist;
> > >> >  }
> > >> >
> > >> >
> > >> This changes also the cfq_should_preempt behaviour, where a large
> > >> seek_mean could be meaningful, so I think it is better to add a
> > >> boolean parameter to cfq_rq_close, to distinguish whether we are
> > >> preempting or looking for queue merges, and make the new code
> > >> conditional on merging.
> > > can you explain why it's meaningful for cfq_should_preempt()? Unless sdist is
> > > very big, for example > 10*seek_mean, the preempt seems not meaningless.
> >
> > Disk access time is a complex function, but for rotational disks it is
> > 'sort of' increasing with the amplitude of the seek. So, if you have a
> > new request that is within the mean seek distance (even if it is
> > larger than our constant), it is good to chose this request instead of
> > waiting for an other one from the active queue (this behaviour is the
> > same exhibited by AS, so we need a good reason to change).
> I have no good reason, but just thought it's a little strange.
> If other ioscheds take the same way, let's stick.
>
>
> seek_mean could be very big sometimes, using it as close criteria is meaningless
> as this doen't improve any performance. So if it's big, let's fallback to
> default value.

Sorry for the lack of response in this thread, christmas vacation is
upon us. The below looks good, I've added it. Thanks!

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/