Re: Re: [patch v2]raid5: fix directio regression

["Jianpeng Ma" <majianpeng@gmail.com>]

On 2012-08-23 14:08 Shaohua Li <shli@kernel.org> Wrote:
>2012/8/16 Shaohua Li <shli@kernel.org>:
>> 2012/8/16 Jianpeng Ma <majianpeng@gmail.com>:
>>> On 2012-08-15 09:44 Shaohua Li <shli@kernel.org> Wrote:
>>>>On Wed, Aug 15, 2012 at 10:56:10AM +1000, NeilBrown wrote:
>>>>> On Tue, 14 Aug 2012 14:33:43 +0800 Shaohua Li <shli@kernel.org> wrote:
>>>>>
>>>>> > On Thu, Aug 09, 2012 at 01:07:01PM +0800, Shaohua Li wrote:
>>>>> > > 2012/8/9 NeilBrown <neilb@suse.de>:
>>>>> > > > On Thu, 9 Aug 2012 09:20:05 +0800 "Jianpeng Ma" <majianpeng@gmail.com> wrote:
>>>>> > > >
>>>>> > > >> On 2012-08-08 20:53 Shaohua Li <shli@kernel.org> Wrote:
>>>>> > > >> >2012/8/8 Jianpeng Ma <majianpeng@gmail.com>:
>>>>> > > >> >> On 2012-08-08 10:58 Shaohua Li <shli@kernel.org> Wrote:
>>>>> > > >> >>>2012/8/7 Jianpeng Ma <majianpeng@gmail.com>:
>>>>> > > >> >>>> On 2012-08-07 13:32 Shaohua Li <shli@kernel.org> Wrote:
>>>>> > > >> >>>>>2012/8/7 Jianpeng Ma <majianpeng@gmail.com>:
>>>>> > > >> >>>>>> On 2012-08-07 11:22 Shaohua Li <shli@kernel.org> Wrote:
>>>>> > > >> >>>>>>>My directIO randomwrite 4k workload shows a 10~20% regression caused by commit
>>>>> > > >> >>>>>>>895e3c5c58a80bb. directIO usually is random IO and if request size isn't big
>>>>> > > >> >>>>>>>(which is the common case), delay handling of the stripe hasn't any advantages.
>>>>> > > >> >>>>>>>For big size request, delay can still reduce IO.
>>>>> > > >> >>>>>>>
>>>>> > > >> >>>>>>>Signed-off-by: Shaohua Li <shli@fusionio.com>
>>>>> > > >> >>>> [snip]
>>>>> > > >> >>>>>>>--
>>>>> > > >> >>>>>> May be used size to judge is not a good method.
>>>>> > > >> >>>>>> I firstly sended this patch, only want to control direct-write-block,not for reqular file.
>>>>> > > >> >>>>>> Because i think if someone used direct-write-block for raid5,he should know the feature of raid5 and he can control
>>>>> > > >> >>>>>> for write to full-write.
>>>>> > > >> >>>>>> But at that time, i did know how to differentiate between regular file and block-device.
>>>>> > > >> >>>>>> I thik we should do something to do this.
>>>>> > > >> >>>>>
>>>>> > > >> >>>>>I don't think it's possible user can control his write to be a
>>>>> > > >> >>>>>full-write even for
>>>>> > > >> >>>>>raw disk IO. Why regular file and block device io matters here?
>>>>> > > >> >>>>>
>>>>> > > >> >>>>>Thanks,
>>>>> > > >> >>>>>Shaohua
>>>>> > > >> >>>> Another problem is the size. How to judge the size is large or not?
>>>>> > > >> >>>> A syscall write is a dio and a dio may be split more bios.
>>>>> > > >> >>>> For my workload, i usualy write chunk-size.
>>>>> > > >> >>>> But your patch is judge by bio-size.
>>>>> > > >> >>>
>>>>> > > >> >>>I'd ignore workload which does sequential directIO, though
>>>>> > > >> >>>your workload is, but I bet no real workloads are. So I'd like
>>>>> > > >> >> Sorry,my explain maybe not corcrect. I write data once which size is almost chunks-size * devices,in order to full-write
>>>>> > > >> >> and as possible as to no pre-read operation.
>>>>> > > >> >>>only to consider big size random directio. I agree the size
>>>>> > > >> >>>judge is arbitrary. I can optimize it to be only consider stripe
>>>>> > > >> >>>which hits two or more disks in one bio, but not sure if it's
>>>>> > > >> >>>worthy doing. Not ware big size directio is common, and even
>>>>> > > >> >>>is, big size request IOPS is low, a bit delay maybe not a big
>>>>> > > >> >>>deal.
>>>>> > > >> >> If add a acc_time for 'striep_head' to control?
>>>>> > > >> >> When get_active_stripe() is ok, update acc_time.
>>>>> > > >> >> For some time, stripe_head did not access and it shold pre-read.
>>>>> > > >> >
>>>>> > > >> >Do you want to add a timer for each stripe? This is even ugly.
>>>>> > > >> >How do you choose the expire time? A time works for harddisk
>>>>> > > >> >definitely will not work for a fast SSD.
>>>>> > > >> A time is like the size which is arbitrary.
>>>>> > > >> How about add a interface in sysfs to control by user?
>>>>> > > >> Only user can judge the workload, which sequatial write or random write.
>>>>> > > >
>>>>> > > > This is getting worse by the minute.  A sysfs interface for this is
>>>>> > > > definitely not a good idea.
>>>>> > > >
>>>>> > > > The REQ_NOIDLE flag is a pretty clear statement that no more requests that
>>>>> > > > merge with this one are expected.  If some use cases sends random requests,
>>>>> > > > maybe it should be setting REQ_NOIDLE.
>>>>> > > >
>>>>> > > > Maybe someone should do some research and find out why WRITE_ODIRECT doesn't
>>>>> > > > include REQ_NOIDLE.  Understanding that would help understand the current
>>>>> > > > problem.
>>>>> > >
>>>>> > > A quick search shows only cfq-iosched uses REQ_NOIDLE. In
>>>>> > > cfq, a queue is idled to avoid losing its share. REQ_NOIDLE
>>>>> > > tells cfq to avoid idle, since the task will not dispatch further
>>>>> > > requests any more. Note this isn't no merge.
>>>>> >
>>>>> > Since REQ_NOIDLE has no relationship with request merge, we'd better remove it.
>>>>> > I came out a new patch, which doesn't depend on request size any more. With
>>>>> > this patch, sequential directio will still introduce unnecessary raid5 preread
>>>>> > (especially for small size IO), but I bet no app does sequential small size
>>>>> > directIO.
>>>>> >
>>>>> > Thanks,
>>>>> > Shaohua
>>>>> >
>>>>> > Subject: raid5: fix directio regression
>>>>> >
>>>>> > My directIO randomwrite 4k workload shows a 10~20% regression caused by commit
>>>>> > 895e3c5c58a80bb. This commit isn't friendly for small size random IO, because
>>>>> > delaying such request hasn't any advantages.
>>>>> >
>>>>> > DirectIO usually is random IO. I thought we can ignore request merge between
>>>>> > bios from different io_submit. So we only consider one bio which can drive
>>>>> > unnecessary preread in raid5, which is large request. If a bio is large enough
>>>>> > and some of its stripes will access two or more disks, such stripes should be
>>>>> > delayed to avoid unnecessary preread till bio for the last disk of the strips
>>>>> > is added.
>>>>> >
>>>>> > REQ_NOIDLE doesn't mean about request merge, I deleted it.
>>>>>
>>>>> Hi,
>>>>>  Have you tested what effect this has on large sequential direct writes?
>>>>>  Because it don't make sense to me and I would be surprised if it improves
>>>>>  things.
>>>>>
>>>>>  You are delaying setting the STRIPE_PREREAD_ACTIVE bit until you think you
>>>>>  have submitted all the writes from this bio that apply to the give stripe.
>>>>>  That does make some sense, however it doesn't seem to deal with the
>>>>>  possibility that the one bio covers parts of two different stripes.  In that
>>>>>  case the first stripe never gets STRIPE_PREREAD_ACTIVE set, so it is delayed
>>>>>  despite having 'REQ_SYNC' set.
>>>>
>>>>I didn't get your point. Isn't last_sector - logical_sector < chunk_sectors true
>>>>in the case?
>>>>
>>>>>  Also, and more significantly, plugging should mean that the various
>>>>>  stripe_heads are not even looked at until all of the original bio is
>>>>>  processed, so while STRIPE_PREREAD_ACTIVE might get set early, it should not
>>>>>  get processed until the whole bio is processed and the queue is unplugged.
>>>>>
>>>>>  So I don't think this patch should make a difference on large direct writes,
>>>>>  and if it does then something strange is going on that I'd like to
>>>>>  understand first.
>>>>
>>>>Aha, ok, this makes sense. recent delayed stripe release should make the
>>>>problem go away. So Jianpeng, can you try your workload with the commit
>>>>reverted with a recent kernel please?
>>>>
>>> I tested used your patch in my workload.
>>> Like the neil said, the performance does not regress.
>>> But if the code is :
>>>>                       if (test_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
>>>>                               release_stripe(sh);
>>>>                       else
>>>>                               release_stripe_plug(mddev, sh);
>>> The speed is about 76MB/s.With those code the speed is 200MB/s.
>>
>> Hmm, what I want to test is upstream kernel with commit 895e3c5c58a80bb
>> reverted. don't apply my patch. We want to just revert the commit.
>
>Did you have data for your original workload with 895e3c5c58a80bb
>reverted now?
our raid5 which had 14 SATA HDDs.

with  895e3c5c58a80bb reverted:
using dd to test 55MB/s
using our-fs 200-250Mb/s

with  895e3c5c58a80bb:
using dd to test 275MB/s
using our-fs 500-550Mb/s