Date:	Thu, 30 Jun 2011 08:29:22 +0800
From:	Shaohua Li <shaohua.li@...el.com>
To:	Vivek Goyal <vgoyal@...hat.com>
Cc:	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	"jaxboe@...ionio.com" <jaxboe@...ionio.com>,
	"linux-fsdevel@...r.kernel.org" <linux-fsdevel@...r.kernel.org>,
	"linux-ext4@...r.kernel.org" <linux-ext4@...r.kernel.org>,
	"khlebnikov@...nvz.org" <khlebnikov@...nvz.org>,
	"jmoyer@...hat.com" <jmoyer@...hat.com>
Subject: Re: [RFC PATCH 0/3] block: Fix fsync slowness with CFQ cgroups

On Wed, Jun 29, 2011 at 09:29:55AM +0800, Vivek Goyal wrote:
> On Wed, Jun 29, 2011 at 09:04:55AM +0800, Shaohua Li wrote:
> 
> [..]
> > > We idle on last queue on sync-noidle tree. So we idle on fysnc queue as
> > > it is last queue on sync-noidle tree. That's how we provide protection
> > > to all sync-noidle queues against sync-idle queues. Instead of idling
> > > on individual quues we do idling in group and that is on service tree.
> > Ok. but this looks silly. We are idling in a noidle service tree or a
> > group (backed by the last queue of the tree or group) because we assume
> > the tree or group can dispatch a request soon. But if the think time of
> > the tree or group is big, the assumption isn't true. Doing idle here is
> > blind. I thought we can extend the think time check for both service
> > tree and group.
> 
> We can implement the thinktime for noidle service tree and group idle as
> well. That's not a problem, though I am yet to be convinced that thinktime
> still makes sense for the group. I guess it will just mean that in the
> past have you done a bunch of IO with gap between IO less than 8ms. If
> yes, then we expect you to do more IO in future. Frankly speaking, I am
> not too sure that how past IO pattern predicts the future IO pattern
> of the group.
> 
> But anyway, the point is, even if you we implement it, it will not solve
> the fsync issue at hand. The reason I explained in previous mail. We 
> will be oscillating between high think time and low thinktime depending
> on whether we are idling or not. There is no correlation between think
> time of fsync thread and idling here.
> 
> I think you are banking on the fact that after fsync, journaling thread
> IO can take more than 8ms hence delaying next IO to fsync thread, pushing
> its thinktim more than 8ms hence we will not idle on fsync thread at
> all. It is just one corner case and I think it is broken in multiple
> cases.
> 
> - If filesystem barriers are disabled or backend storage has battery
>   backup then journal IO most likely will go in cache and barriers
>   will be ignored. In that case write will finish almost instantly
>   and we will get next IO from fsync thread very soon hence pushing
>   down thinktime of fsync thread which will enable idling and we will
>   be back to the problem we are trying to solve.
> 
> - Fsync thread might be submitting string of IOs (say 10-12) before it
>   moves to journal thread to commit meta data. In that case we might
>   have lowered thinktime of fsync hence enable idle. 
> 
> So implementing think time for service tree/group might be a good idea
> in general but it will not solve this IO dependecny issue across cgroups.
Ok, fair enough. I'll give a try and check how things change with the fsync workload.

Thanks,
Shaohua
--
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