Date:	Fri, 7 Aug 2009 10:50:04 +0200
From:	Jens Axboe <jens.axboe@...cle.com>
To:	Jeff Garzik <jeff@...zik.org>
Cc:	Alan Cox <alan@...rguk.ukuu.org.uk>, linux-kernel@...r.kernel.org,
	linux-scsi@...r.kernel.org, Eric.Moore@....com
Subject: Re: [PATCH 1/3] block: add blk-iopoll, a NAPI like approach for
	block  devices

On Fri, Aug 07 2009, Jeff Garzik wrote:
> Jens Axboe wrote:
>> On Thu, Aug 06 2009, Alan Cox wrote:
>>>> doing the command completion when the irq occurs, schedule a dedicated
>>>> softirq in the hopes that we will complete more IO when the iopoll
>>>> handler is invoked. Devices have a budget of commands assigned, and will
>>>> stay in polled mode as long as they continue to consume their budget
>>>> from the iopoll softirq handler. If they do not, the device is set back
>>>> to interrupt completion mode.
>>> This seems a little odd for pure ATA except for NCQ commands. Normal ATA
>>> is notoriously completion/reissue latency sensitive [to the point I
>>> suspect we should be dequeuing 2 commands from SCSI and loading the next
>>> in the completion handler as soon as we recover the result task file and
>>> see no error rather than going up and down the stack)
>>
>> Yes certainly, it's only for devices that do queuing. If they don't,
>> then we will always have just the one command to complete. So not much
>> to poll! As to pre-prep for extra latency intensive devices, have you
>> tried experimenting with just pretending that non-ncq devices in libata
>> have a queue depth of 2? That should ensure that the first command
>> available upon completion of the existing command is already prepped.
>> Not sure how much time that would save, I would hope that our prep phase
>> isn't too slow to begin with (or that would be the place to fix :-)
>>
>>> What do the numbers look like ?
>>
>> On a slow box (with many cores), the benefits are quite huge:
>>
>>
>> blocksize       blk-iopoll      IOPS    IRQ/sec         Commands/IRQ
>> --------------------------------------------------------------------
>> 512b            0               25168   ~19500          1,3
>> 512b            1               30355     ~750          40
>> 4096b           0               25612   ~21500          1,2
>> 4096b           1               30231    ~1200          25
>>
>> I suspect there's some cache interaction going on here too, but the
>> numbers do look very good. On a faster box (and different architecture),
>> on a test that does 50k IOPS, they perform identically but the iopoll
>> approach uses less CPU. The interrupt rate drops from 55k ints/sec to
>> 39-40k ints/sec for that case.
>
> It's easy to move work from one place to another, so I would definitely  
> expect that IRQ/sec drops...  but these are the more relevant numbers, 
> IMO:
>
> * CPU usage before/after
> * latency before/after

As I mentioned in the 0/3 email, latency for my tests were as good or
better than the original an CPU usage was lower. The former must largely
be due to decreased latency in commands successfully retired in addition
to the one that triggered the IRQ, since the latency for the first
command should be a little higher. Since we use softirq completion for
the command in the FIRST place anyway, it probably wont make any
difference (and this latency for the first command should be almost
immeasurably from the non-iopoll path).

> Also, and even for storage where command queueing is _possible_, there  
> is a problem case we saw with NAPI:  sometimes the combination of a fast  
> computer and an under-100%-utilization workload can imply repeated cycles 
> of
>
> 	spin lock
> 	irq disable
> 	blk_iopoll_sched()
> 	spin unlock
>
> 	spin lock
> 	handle a single command completion
> 	spin unlock
> 	blk_iopoll_complete()
>
> which not only erases the benefit, but winds up being more costly, both  
> in terms of CPU usage and in terms of latency.

It's clear that if you always only retire a single command AND you need
to lock at both ends, then it'll never be a win. I guess we could detect
such cases and be more cautious about when to enter iopoll, if that is
an issue. The ahci case looks like what you describe and I'm not seeing
any issues on the laptop, but I do concede that this is something to
look out for. If you look at the mpt conversion, we don't get cache line
bouncing on a lock there. As I also wrote, ahci is only really
interesting for test purposes, I don't envision a lot of real world win
there. But it widens the scope for testing :-)

> This makes measuring the problem much more difficult; the interesting  
> case I am highlighting does not occur when using a benchmarking tool to  
> keep a storage device at 100% utilization.

Of course not, that case is primarily interesting to gauge potential
best case wins.

> We don't want to optimize for the 100%-load case at the expense of the  
> _common case_, which is IMO utilization below 100%.  Servers are not  
> 100% busy all the time, which opens the possibility that a  
> split-completion scheme such as the one presented can actually use  
> _more_ CPU than the current, unmodified 2.6.31-rc kernel.

Depends, if the common case doesn't really suffer, then it doesn't
matter. Graceful load handling is important.

> I'm not NAK'ing...  just inserting some relevant NAPI field experience,  
> and hoping for some numbers that better measure the costs/benefits.

Appreciate you looking over this, and I'll certainly be posting some
more numbers on this. It'll largely depend on both storage, controller,
and worload.

-- 
Jens Axboe

--
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