Message-ID: <20091116144341.GC13235@redhat.com>
Date:	Mon, 16 Nov 2009 09:43:41 -0500
From:	Vivek Goyal <vgoyal@...hat.com>
To:	Corrado Zoccolo <czoccolo@...il.com>
Cc:	Linux-Kernel <linux-kernel@...r.kernel.org>,
	Jens Axboe <jens.axboe@...cle.com>,
	Jeff Moyer <jmoyer@...hat.com>
Subject: Re: [PATCH] cfq-iosched: fix ncq detection code

On Sat, Nov 14, 2009 at 09:01:45PM +0100, Corrado Zoccolo wrote:
> Hi Vivek,
> On Sat, Nov 14, 2009 at 8:09 PM, Vivek Goyal <vgoyal@...hat.com> wrote:
> > On Sat, Nov 14, 2009 at 02:33:40PM +0100, Corrado Zoccolo wrote:
> >> CFQ's detection of queueing devices assumes a non-queuing device and detects
> >> if the queue depth reaches a certain threshold.
> >
> > Hi Corrado,
> >
> > In cfq_init_queue(), we init hw_tag = 1. So I think we start with the
> > assumption that device supports NCQ.
> Yes, it is a cut&paste typo. It should have been:
> CFQ's detection of queueing devices assumes a queuing device and detects
> 
> >
> >>  Under some workloads (e.g.
> >> synchronous reads), CFQ effectively forces a unit queue depth, thus defeating
> >> the detection logic.  This leads to poor performance on queuing hardware,
> >> since the idle window remains enabled.
> >
> > Is it always the case that NCQ hardware performs better if we don't idle.
> > I got a rotational disk which supports NCQ but if we don't idle, than it
> > becomes seek bound and throughput drops.
> 
> I think this is not the main point. The real point of this patch is to
> detect NCQ correctly.
> After the detection is correct, we can have a sound policy that
> decides when and how much to idle.
> For the how-much part, I'm working on measuring the cost of a seek, so
> we can wait proportionally for that cost.
> This will imply a reasonable idle for rotational (even NCQ) media, and
> small or no idle for NCQ SSD.

Ok, idling based on some kind of feedback mechanism based on seek cost
makes sense.

So this patch is in preparation of your seek cost feedback mechanism. The
only affect of this currently seem to be that we will not arm the idle
timer on NCQ SSD. Which is good I think.

Actually, even if we not arm the timer, just keeping slice_idle non zero
is hurting on SSD. If slice_idle is non zero, we are still driving queue
depth as 1 and not achieving the full potential of SSD.

If slice_idle=0, everything belongs to noidle group as soon as new queue gets
backlogged, it preempts the current queue and we dispatch new request
hence driving deeper queue depths and achieving more out of SSD.

Driving smaller queue depth hurts more if I am doing direct IO in smaller
block sizes. Launched 8 processes doing direct IO in 4K size blocks and
I get 20MB/s. Same workload launched with slice_idle=0 and I get 100MB/s.
Five times better throughput. I am wondering, if there is any point in 
keeping slice_idle non zero on NCQ SSDs.

Thanks
Vivek
 
> 
> > On SSD also, seeks are cheap but not free. So if workload is less, then
> > we might prefer to idle and I think that's the purpose of this dynamic
> > hw_tag thing. If non-idling kind of workload is sufficient, then disable
> > idling otherwise in case of sync workloads, continue to idle.
> I don't think so. The dynamic hw_tag purpose was to determine if for
> some reason the depth was dynamically altered, but it doesn't work
> very well, so we can't rely on it.
> 
> >
> > In fact, whether device supports NCQ or not, isn't this information
> > statically available? I do a cat on "/sys/block/<sdd>/device/queue_depth"
> > and on non queuing hardware it has been set to 1 and on queuing hardware
> > it has been set to 31 by default after system boot. So I am assuming we
> > statically have the information about device queuing capability? At least
> > for the locally connected devices.
> It is only for devices that plug into the SCSI subsystem.
> >>
> >> Given this premise, switching to hw_tag = 0 after we have proved at
> >> least once that the device is NCQ capable is not a good choice.
> >
> > If it is proven that NCQ always means that idling is not good.
> No. The hidden assumption is 'if we trust in our policy'.
> If the policy is not good, maybe it is because, due to the fluctuating
> NCQ detection, we could never test it thoroughly.
> > Not true
> > for atleast locally connected media. But is probably true many a times
> > for SSD and fast storage arrays.
> 
> Thanks,
> Corrado
> >
> > Thanks
> > Vivek
> >
> >>
> >> The new detection code starts in an indeterminate state, in which CFQ behaves
> >> as if hw_tag = 1, and then, if for a long observation period we never saw
> >> large depth, we switch to hw_tag = 0, otherwise we stick to hw_tag = 1,
> >> without reconsidering it again.
> >>
> >> Signed-off-by: Corrado Zoccolo <czoccolo@...il.com>
> >> ---
> >>  block/cfq-iosched.c |   24 +++++++++++++++---------
> >>  1 files changed, 15 insertions(+), 9 deletions(-)
> >>
> >> diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> >> index 1bcbd8c..6925ab9 100644
> >> --- a/block/cfq-iosched.c
> >> +++ b/block/cfq-iosched.c
> >> @@ -191,8 +191,14 @@ struct cfq_data {
> >>        */
> >>       int rq_queued;
> >>       int hw_tag;
> >> -     int hw_tag_samples;
> >> -     int rq_in_driver_peak;
> >> +     /*
> >> +      * hw_tag can be
> >> +      * -1 => indeterminate, (cfq will behave as if NCQ is present, to allow better detection)
> >> +      *  1 => NCQ is present (hw_tag_est_depth is the estimated max depth)
> >> +      *  0 => no NCQ
> >> +      */
> >> +     int hw_tag_est_depth;
> >> +     unsigned int hw_tag_samples;
> >>
> >>       /*
> >>        * idle window management
> >> @@ -2527,8 +2533,11 @@ static void cfq_update_hw_tag(struct cfq_data *cfqd)
> >>  {
> >>       struct cfq_queue *cfqq = cfqd->active_queue;
> >>
> >> -     if (rq_in_driver(cfqd) > cfqd->rq_in_driver_peak)
> >> -             cfqd->rq_in_driver_peak = rq_in_driver(cfqd);
> >> +     if (rq_in_driver(cfqd) > cfqd->hw_tag_est_depth)
> >> +             cfqd->hw_tag_est_depth = rq_in_driver(cfqd);
> >> +
> >> +     if (cfqd->hw_tag == 1)
> >> +             return;
> >>
> >>       if (cfqd->rq_queued <= CFQ_HW_QUEUE_MIN &&
> >>           rq_in_driver(cfqd) <= CFQ_HW_QUEUE_MIN)
> >> @@ -2547,13 +2556,10 @@ static void cfq_update_hw_tag(struct cfq_data *cfqd)
> >>       if (cfqd->hw_tag_samples++ < 50)
> >>               return;
> >>
> >> -     if (cfqd->rq_in_driver_peak >= CFQ_HW_QUEUE_MIN)
> >> +     if (cfqd->hw_tag_est_depth >= CFQ_HW_QUEUE_MIN)
> >>               cfqd->hw_tag = 1;
> >>       else
> >>               cfqd->hw_tag = 0;
> >> -
> >> -     cfqd->hw_tag_samples = 0;
> >> -     cfqd->rq_in_driver_peak = 0;
> >>  }
> >>
> >>  static void cfq_completed_request(struct request_queue *q, struct request *rq)
> >> @@ -2960,7 +2966,7 @@ static void *cfq_init_queue(struct request_queue *q)
> >>       cfqd->cfq_slice_async_rq = cfq_slice_async_rq;
> >>       cfqd->cfq_slice_idle = cfq_slice_idle;
> >>       cfqd->cfq_latency = 1;
> >> -     cfqd->hw_tag = 1;
> >> +     cfqd->hw_tag = -1;
> >>       cfqd->last_end_sync_rq = jiffies;
> >>       return cfqd;
> >>  }
> >> --
> >> 1.6.2.5
> >>
> >
> 
> 
> 
> -- 
> __________________________________________________________________________
> 
> dott. Corrado Zoccolo                          mailto:czoccolo@...il.com
> PhD - Department of Computer Science - University of Pisa, Italy
> --------------------------------------------------------------------------
> The self-confidence of a warrior is not the self-confidence of the average
> man. The average man seeks certainty in the eyes of the onlooker and calls
> that self-confidence. The warrior seeks impeccability in his own eyes and
> calls that humbleness.
>                                Tales of Power - C. Castaneda
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists