Re: Bad SSD performance with recent kernels

From: Wu Fengguang <wfg@linux.intel.com>
To: Shaohua Li <shaohua.li@intel.com>
Cc: Vivek Goyal <vgoyal@redhat.com>,
	Eric Dumazet <eric.dumazet@gmail.com>,
	Herbert Poetzl <herbert@13thfloor.at>,
	Andrew Morton <akpm@linux-foundation.org>,
	LKML <linux-kernel@vger.kernel.org>, Jens Axboe <axboe@kernel.dk>,
	Tejun Heo <tj@kernel.org>
Subject: Re: Bad SSD performance with recent kernels
Date: Tue, 31 Jan 2012 09:07:55 +0800	[thread overview]
Message-ID: <20120131010755.GA12776@localhost> (raw)
In-Reply-To: <1327968859.21268.12.camel@sli10-conroe>

On Tue, Jan 31, 2012 at 08:14:19AM +0800, Li, Shaohua wrote:
> On Mon, 2012-01-30 at 17:26 -0500, Vivek Goyal wrote:
> > On Mon, Jan 30, 2012 at 03:51:49PM +0100, Eric Dumazet wrote:
> > > Le lundi 30 janvier 2012 à 22:28 +0800, Wu Fengguang a écrit :
> > > > On Mon, Jan 30, 2012 at 06:31:34PM +0800, Li, Shaohua wrote:
> > > > 
> > > > > Looks the 2.6.39 block plug introduces some latency here. deleting
> > > > > blk_start_plug/blk_finish_plug in generic_file_aio_read seems
> > > > > workaround
> > > > > the issue. The plug seems not good for sequential IO, because readahead
> > > > > code already has plug and has fine grained control.
> > > > 
> > > > Why not remove the generic_file_aio_read() plug completely? It
> > > > actually prevents unplugging immediately after the readahead IO is
> > > > submitted and in turn stalls the IO pipeline as showed by Eric's
> > > > blktrace data.
> > > > 
> > > > Eric, will you test this patch? Thank you.
> > 
> > Can you please run the blktrace again with this patch applied. I am curious
> > to see how does traffic pattern look like now.
> > 
> > In your previous trace, there were so many small 8 sector requests which
> > were merged into 512 sector requests before dispatching to disk. (I am
> > not sure why those requests are not bigger. Shouldn't readahead logic
> > submit a bigger request?) Now with plug/unplug logic removed, I am assuming
> > we should be doing less merging and dispatching more smaller requests. May be
> > that is helping and cutting down on disk idling time.
> > 
> > In previous logs, 512 sector request seems to be taking around 1ms to
> > complete after dispatch. In between requests disk seems to be idle
> > for around .5 to .6 ms. Out of this .3 ms seems to be gone in just
> > coming up with new request after completion of previous one and another
> > .3ms seems to be consumed in merging the smaller IOs. So if we don't wait
> > for merging, it should keep disk busier for .3ms more which is 30% of time
> > it takes to complete 512 sector request. So theoritically it can give
> > 30% boost for this workload. (Assuming request size will not impact the
> > disk throughput very severely).
> > 
> > Anyway, some blktrace data will shed some light..
> yep, I suspect plug merges big request too (iostat shows it too), that's
> why I only think delete the plug in generic_file_aio_read as a
> workaround.

It's good to merge requests inside the same readahead window. However
I don't think readahead window A should be merged with B at the cost
of delaying A for some time, which will break the pipeline. If larger
IO is desirable, we can do so by increasing the readahead size.

> I still thought readahead has something to do here. I
> observed the async readahead does readahead (A, A + 2M), and follows (A
> +128k, A+2M), (A+256k, A+2M) ..., the later readahead doesn't work
> because we already have (A, A+2M) in memory at that time. Anyway, I can
> reproduce the issue, will play with it more today.

How do you observe that? I don't think that readahead pattern is
possible. However I do see such _read_ patterns.

Thanks,
Fengguang