Re: [RFC] block/throttle: Add IO throttled information in blkcg.

[Vivek Goyal <vgoyal@redhat.com>]

On Tue, May 22, 2012 at 11:14:55PM +0800, Tao Ma wrote:
> On 05/22/2012 11:06 PM, Vivek Goyal wrote:
> > On Tue, May 22, 2012 at 10:44:11PM +0800, Tao Ma wrote:
> >> Hi Vivek,
> >> 	Thanks for the quick response.
> >> On 05/22/2012 07:11 PM, Vivek Goyal wrote:
> >>> On Tue, May 22, 2012 at 04:10:36PM +0800, Tao Ma wrote:
> >>>> From: Tao Ma <boyu.mt@taobao.com>
> >>>>
> >>>> Currently, if the IO is throttled by io-throttle, the SA has no idea of
> >>>> the situation and can't report it to the real application user about
> >>>> that he/she has to do something. So this patch adds a new interface
> >>>> named blkio.throttle.io_throttled which indicates how many IOs are
> >>>> currently throttled.
> >>>
> >>> If the only purpose is to know whether IOs are being throttled, why
> >>> not just scan for the rules and see if respective device has any
> >>> throttling rules or not.
> >> Sorry, but setting a throttling rules doesn't mean the IOs are
> >> throttled, right? So scanning doesn't work here IMHO.
> > 
> > It means IOs will be throttled if you cross a certain rate. But yes, it
> > does not give any information that if at time T if there are any bios
> > throttled in the queue or not.
> > 
> >>>
> >>> Even if you introduce this interface, you will end up scanning for
> >>> throttled ios against that particular device. And if IO is not happening
> >>> at that moment or if IO rate is not exceeding the rate limit, there
> >>> might not be any throttled ios and one might get misled.
> >> Oh, no actually in a *clound computing* environment, it is really
> >> useful, not misled. So let me describe it in more detail. Our product
> >> system will limit every instance to an approximate number at first, and
> >> then watch out the IOs being throttled. If these numbers is high, it can:
> >> 1) Shout loudly to the application programmer about the abuse if he
> >> sends out too much IO requests.
> >> 2) If it is not too much and some other instances are not active, adjust
> >> the throttled ratio so that this instance can work much faster.
> > 
> > Ok, so you want to use this more as "congestion" parameter which tells at
> > a given moment how busy the queue is, or in this instance how many IOs
> > are backlogged in a cgroup due to throttling limits.
> yeah, with this information the daemon can adjust these limits
> automatically.

I am hoping that this daemon will monitor the file for long periods and
will not reach to bursty traffic from application.

> > 
> > I guess, it is not a bad idea to export this stat then. Will
> > "blkio.throttle.queued" be a better name to reflect that how many bios
> > are currently queued in throttling layer of request queue.
> I have thought of this name at the very first time. But there is also
> another one named "blkio.queued" which indicated the IOs being queued in
> the scheduler. I don't want the user to be confused and that's the
> reason I use "blkio.throttle.io_throttled".

Actually it is blkio.io_queued which shows number of requests queued in
CFQ in that cgroup.

CFQ and throttling are two different policies and they have separate
files in cgroup. Ideally blkio.io_queued should have been
blkio.cfq.io_queued but initially it did not occur to me that I should
qualify these files with policy name.

Later when throttling policy came along, then I qualified new files with
policy name. blkio.throttle.*.

In summary, blkio.io_queued gives stats of io queued at CFQ level. So it
makes sense to create blkio.throttle.io_queued which tells how many
bios are currently throttled and queued in throttling layer in this
request queue from this cgroup.

Thanks
Vivek