Message-ID: <20120116082927.GF3174@redhat.com>
Date:	Mon, 16 Jan 2012 03:29:27 -0500
From:	Vivek Goyal <vgoyal@...hat.com>
To:	Shaohua Li <shaohua.li@...el.com>
Cc:	Dave Chinner <david@...morbit.com>, linux-kernel@...r.kernel.org,
	axboe@...nel.dk, jmoyer@...hat.com
Subject: Re: [RFC 0/3]block: An IOPS based ioscheduler

On Mon, Jan 16, 2012 at 03:55:41PM +0800, Shaohua Li wrote:
> On Mon, 2012-01-16 at 02:11 -0500, Vivek Goyal wrote:
> > On Mon, Jan 16, 2012 at 12:36:30PM +0800, Shaohua Li wrote:
> > > On Sun, 2012-01-15 at 17:45 -0500, Vivek Goyal wrote:
> > > > On Mon, Jan 09, 2012 at 09:09:35AM +0800, Shaohua Li wrote:
> > > > 
> > > > [..]
> > > > > > You need to present raw numbers and give us some idea of how close
> > > > > > those numbers are to raw hardware capability for us to have any idea
> > > > > > what improvements these numbers actually demonstrate.
> > > > > Yes, your guess is right. The hardware has limitation. 12 SSD exceeds
> > > > > the jbod capability, for both throughput and IOPS, that's why only
> > > > > read/write mixed workload impacts. I'll use less SSD in later tests,
> > > > > which will demonstrate the performance better. I'll report both raw
> > > > > numbers and fiops/cfq numbers later.
> > > > 
> > > > If fiops number are better please explain why those numbers are better.
> > > > If you cut down on idling, it is obivious that you will get higher
> > > > throughput on these flash devices. CFQ does disable queue idling for
> > > > non rotational NCQ devices. If higher throughput is due to driving
> > > > deeper queue depths, then CFQ can do that too just by changing quantum
> > > > and disabling idling. 
> > > it's because of quantum. Surely you can change the quantum, and CFQ
> > > performance will increase, but you will find CFQ is very unfair then.
> > 
> > Why increasing quantum leads to CFQ being unfair? In terms of time it
> > still tries to be fair. 
> we can dispatch a lot of requests to NCQ SSD with very small time
> interval. The disk can finish a lot of requests in small time interval
> too. The time is much smaller than 1 jiffy. Increasing quantum can lead
> a task dispatches request more faster and makes the accounting worse,
> because with small quantum the task needs wait to dispatch. you can
> easily verify this with a simple fio test.
> 
> > That's a different thing that with NCQ, right
> > time measurement is not possible with requests from multiple queues
> > being in the driver/disk at the same time. So accouting in terms of
> > iops per queue might make sense.
> yes.
> 
> > > > So I really don't understand that what are you doing fundamentally
> > > > different in FIOPS ioscheduler. 
> > > > 
> > > > The only thing I can think of more accurate accounting per queue in
> > > > terms of number of IOs instead of time. Which can just serve to improve
> > > > fairness a bit for certain workloads. In practice, I think it might
> > > > not matter much.
> > > If quantum is big, CFQ will have better performance, but it actually
> > > fallbacks to Noop, no any fairness. fairness is important and is why we
> > > introduce CFQ.
> > 
> > It is not exactly noop. It still preempts writes and prioritizes reads
> > and direct writes. 
> sure, I mean fairness mostly here.
> 
> > Also, what's the real life workload where you face issues with using
> > say deadline with these flash based storage.
> deadline doesn't provide fairness. mainly cgroup workload. workload with
> different ioprio has issues too, but I don't know which real workload
> uses ioprio.

Personally I have not run into any workload which provides deep queue depths
constantly for a very long time. I had to run fio to create such
scnearios.

Not running deep queue depths will lead to expiration of queue (Otherwise
idling will kill performance on these fast devices). And without idling
most of the logic of slice and accounting does not help. A queue
dispatches some requests and expires (irrespective of what time slice
you had allocated it based on ioprio).

That's why I am insisting that it would be nice that any move in this
direction should be driven by some real workload instead of just coming
up with synthetic workloads.

> 
> > > 
> > > In summary, CFQ isn't both fair and good performance. FIOPS is trying to
> > > be fair and have good performance. I didn't think any time based
> > > accounting can make the goal happen for NCQ and SSD (even cfq cgroup
> > > code has iops mode, so suppose you should already know this well).
> > > 
> > > Surely you can change CFQ to make it IOPS based, but this will mess the
> > > code a lot, and FIOPS shares a lot of code with CFQ. So I'd like to have
> > > a separate ioscheduler which is IOPS based.
> > 
> > I think writing a separate IO scheduler just to do accouting in IOPS while
> > retaining rest of the CFQ code is not a very good idea. Modifying CFQ code
> > to be able to deal with both time based as well as IOPS accounting might
> > turn out to be simpler.
> changing CFQ works, but I really want to avoid having something like
> if (iops) 
>    xxx
> else
>    xxx

I think you can provide a wrapper function and abstract out unit of time
as "charge" or something like that. 

> I plan adding scales for read/write, request size, etc, because
> read/write cost is different and request with different size has
> different cost in SSD. This can be added to CFQ too with pain. That said
> I didn't completely object to make CFQ support IOPS accounting, but my
> feeling is a separate ioscheduler is more clean.

You can provide one function to calculate the "charge" which can be either
time or some kind of IOPS unit adjusted with request size and use that.

Given the fact that new IOPS scheduler will be sharing lots of code with
CFQ (I presume 80-90% of concepts are same) and the only major difference
is accounting, I would tend to think that modifying CFQ is better.

Jens, what do you think?

Thanks
Vivek
--
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