Date: Sun, 19 Apr 2009 08:59:03 -0400
From: Vivek Goyal <vgoyal@redhat.com>
To: Nauman Rafique <nauman@google.com>
Cc: Andrea Righi <righi.andrea@gmail.com>,
       Andrew Morton <akpm@linux-foundation.org>, dpshah@google.com,
       lizf@cn.fujitsu.com, mikew@google.com, fchecconi@gmail.com,
       paolo.valente@unimore.it, jens.axboe@oracle.com, ryov@valinux.co.jp,
       fernando@intellilink.co.jp, s-uchida@ap.jp.nec.com, taka@valinux.co.jp,
       guijianfeng@cn.fujitsu.com, arozansk@redhat.com, jmoyer@redhat.com,
       oz-kernel@redhat.com, dhaval@linux.vnet.ibm.com,
       balbir@linux.vnet.ibm.com, linux-kernel@vger.kernel.org,
       containers@lists.linux-foundation.org, menage@google.com,
       peterz@infradead.org
Subject: Re: IO controller discussion (Was: Re: [PATCH 01/10] Documentation)
Message-ID: <20090419125903.GB8493@redhat.com>
References: <1236823015-4183-2-git-send-email-vgoyal@redhat.com> <20090312001146.74591b9d.akpm@linux-foundation.org> <20090312180126.GI10919@redhat.com> <49D8CB17.7040501@gmail.com> <20090407064046.GB20498@redhat.com> <20090408203756.GB10077@linux> <20090416183753.GE8896@redhat.com> <20090417093656.GA5246@linux> <20090417141358.GD29086@redhat.com> <e98e18940904171109r17ccb054kb7879f8d02ac26b5@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <e98e18940904171109r17ccb054kb7879f8d02ac26b5@mail.gmail.com>
User-Agent: Mutt/1.5.18 (2008-05-17)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 9661
Lines: 187

On Fri, Apr 17, 2009 at 11:09:51AM -0700, Nauman Rafique wrote:
> On Fri, Apr 17, 2009 at 7:13 AM, Vivek Goyal <vgoyal@redhat.com> wrote:
> > On Fri, Apr 17, 2009 at 11:37:28AM +0200, Andrea Righi wrote:
> >> On Thu, Apr 16, 2009 at 02:37:53PM -0400, Vivek Goyal wrote:
> >> > > I think it would be possible to implement both proportional and limiting
> >> > > rules at the same level (e.g., the IO scheduler), but we need also to
> >> > > address the memory consumption problem (I still need to review your
> >> > > patchset in details and I'm going to test it soon :), so I don't know if
> >> > > you already addressed this issue).
> >> > >
> >> >
> >> > Can you please elaborate a bit on this? Are you concerned about that data
> >> > structures created to solve the problem consume a lot of memory?
> >>
> >> Sorry I was not very clear here. With memory consumption I mean wasting
> >> the memory with hard/slow reclaimable dirty pages or pending IO
> >> requests.
> >>
> >> If there's only a global limit on dirty pages, any cgroup can exhaust
> >> that limit and cause other cgroups/processes to block when they try to
> >> write to disk.
> >>
> >> But, ok, the IO controller is not probably the best place to implement
> >> such functionality. I should rework on the per cgroup dirty_ratio:
> >>
> >> https://lists.linux-foundation.org/pipermail/containers/2008-September/013140.html
> >>
> >> Last time we focused too much on the best interfaces to define dirty
> >> pages limit, and I never re-posted an updated version of this patchset.
> >> Now I think we can simply provide the same dirty_ratio/dirty_bytes
> >> interface that we provide globally, but per cgroup.
> >>
> >> >
> >> > > IOW if we simply don't dispatch requests and we don't throttle the tasks
> >> > > in the cgroup that exceeds its limit, how do we avoid the waste of
> >> > > memory due to the succeeding IO requests and the increasingly dirty
> >> > > pages in the page cache (that are also hard to reclaim)? I may be wrong,
> >> > > but I think we talked about this problem in a previous email... sorry I
> >> > > don't find the discussion in my mail archives.
> >> > >
> >> > > IMHO a nice approach would be to measure IO consumption at the IO
> >> > > scheduler level, and control IO applying proportional weights / absolute
> >> > > limits _both_ at the IO scheduler / elevator level _and_ at the same
> >> > > time block the tasks from dirtying memory that will generate additional
> >> > > IO requests.
> >> > >
> >> > > Anyway, there's no need to provide this with a single IO controller, we
> >> > > could split the problem in two parts: 1) provide a proportional /
> >> > > absolute IO controller in the IO schedulers and 2) allow to set, for
> >> > > example, a maximum limit of dirty pages for each cgroup.
> >> > >
> >> >
> >> > I think setting a maximum limit on dirty pages is an interesting thought.
> >> > It sounds like as if memory controller can handle it?
> >>
> >> Exactly, the same above.
> >
> > Thinking more about it. Memory controller can probably enforce the higher
> > limit but it would not easily translate into a fixed upper async write
> > rate. Till the process hits the page cache limit or is slowed down by
> > dirty page writeout, it can get a very high async write BW.
> >
> > So memory controller page cache limit will help but it would not direclty
> > translate into what max bw limit patches are doing.
> >
> > Even if we do max bw control at IO scheduler level, async writes are
> > problematic again. IO controller will not be able to throttle the process
> > until it sees actuall write request. In big memory systems, writeout might
> > not happen for some time and till then it will see a high throughput.
> >
> > So doing async write throttling at higher layer and not at IO scheduler
> > layer gives us the opprotunity to produce more accurate results.
> 
> Wouldn't 'doing control on writes at a higher layer' have the same
> problems as the ones we talk about in dm-ioband? What if the cgroup
> being throttled for dirtying pages has a high weight assigned to it at
> the IO scheduler level? What if there are threads of different classes
> within that cgroup, and we would want to let RT task dirty the pages
> before BE tasks? I am not sure all these questions make sense, but
> just wanted to raise issues that might pop up.
> 

Yes, i would think that you will run into same issues if one is not
maintaining separate queues. The only difference is that we are throttling
only async writes and i am not sure how well the notion of fairness is
defined for async writes in current system as most of the time nobody
has cared too much for latency of async writes (fsync is the exception).

But in general, yes, putting things into a single queue again puts into
the same situation where we loose the notion of class and prio with-in
cgroup. So any design which tries to do that probably is not a very good
idea.

> If the whole system is designed with cgroups in mind, then throttling
> at IO scheduler layer should lead to backlog, that could be seen at
> higher level. For example, if a cgroup is not getting service at IO
> scheduler level, it should run out of request descriptors, and thus
> the thread writing back dirty pages should notice it (if its pdflush,
> blocking it is probably not the best idea). And that should mean the
> cgroup should hit the dirty threshold, and disallow the task to dirty
> further pages. There is a possibility though that getting all this
> right might be an overkill and we can get away with a simpler
> solution. One possibility seems to be that we provide some feedback
> from IO scheduling layer to higher layers, that cgroup is hitting its
> write bandwith limit, and should not be allowed to dirty any more
> pages.

I think bdi_write_congested(), already provides the feedback to upper
layer about how congested the queue is and it is likely that any thread
submitting new write request to this bdi will be made to sleep before
it gets the reqeust descriptor.

I think in new scheme of thing, this notion of per device congestion shall
have to changed into per device per cgroup congestion.

Thanks
Vivek


> 
> >
> > For sync requests, I think IO scheduler max bw control should work fine.
> >
> > BTW, andrea, what is the use case of your patches? Andrew had mentioned
> > that some people are already using it. I am curious to know will a
> > proportional BW controller will solve the issues/requirements of these
> > people or they have specific requirement of traffic shaping and max bw
> > controller only.
> >
> > [..]
> >> > > > Can you please give little more details here regarding how QoS requirements
> >> > > > are not met with proportional weight?
> >> > >
> >> > > With proportional weights the whole bandwidth is allocated if no one
> >> > > else is using it. When IO is submitted other tasks with a higher weight
> >> > > can be forced to sleep until the IO generated by the low weight tasks is
> >> > > not completely dispatched. Or any extent of the priority inversion
> >> > > problems.
> >> >
> >> > Hmm..., I am not very sure here. When admin is allocating the weights, he
> >> > has the whole picture. He knows how many groups are conteding for the disk
> >> > and what could be the worst case scenario. So if I have got two groups
> >> > with A and B with weight 1 and 2 and both are contending, then as an
> >> > admin one would expect to get 33% of BW for group A in worst case (if
> >> > group B is continuously backlogged). If B is not contending than A can get
> >> > 100% of BW. So while configuring the system, will one not plan for worst
> >> > case (33% for A, and 66 % for B)?
> >>
> >> OK, I'm quite convinced.. :)
> >>
> >> To a large degree, if we want to provide a BW reservation strategy we
> >> must provide an interface that allows cgroups to ask for time slices
> >> such as max/min 5 IO requests every 50ms or something like that.
> >> Probably the same functionality can be achieved translating time slices
> >> from weights, percentages or absolute BW limits.
> >
> > Ok, I would like to split it in two parts.
> >
> > I think providng minimum gurantee in absolute terms like 5 IO request
> > every 50ms will be very hard because IO scheduler has no control over
> > how many competitors are there. An easier thing will be to have minimum
> > gurantees on share basis. For minimum BW (disk time slice) gurantee, admin
> > shall have to create right cgroup hierarchy and assign weights properly and
> > then admin can calculate what % of disk slice a particular group will get
> > as minimum gurantee. (This is more complicated than this as there are
> > time slices which are not accounted to any groups. During queue switch
> > cfq starts the time slice counting only after first request has completed
> > to offset the impact of seeking and i guess also NCQ).
> >
> > I think it should be possible to give max bandwidth gurantees in absolute
> > terms, like io/s or sectors/sec or MB/sec etc, because only thing IO
> > scheduler has to do is to not allow dispatch from a particular queue if
> > it has crossed its limit and then either let the disk idle or move onto
> > next eligible queue.
> >
> > The only issue here will be async writes. max bw gurantee for async writes
> > at IO scheduler level might not mean much to application because of page
> > cache.
> >
> > Thanks
> > Vivek
> >
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/