On Mon, Sep 14 2020 at 9:33pm -0400,
Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
> On Thu, Sep 10 2020 at 3:29pm -0400,
> Vijayendra Suman <vijayendra.suman@xxxxxxxxxx> wrote:
>
> > Hello Mike,
> >
> > I checked with upstream, performance measurement is similar and
> > shows performance improvement when
> > 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74 is reverted.
> >
> > On 9/10/2020 7:54 PM, Mike Snitzer wrote:
> > >[cc'ing dm-devel and linux-block because this is upstream concern too]
> > >
> > >On Wed, Sep 09 2020 at 1:00pm -0400,
> > >Vijayendra Suman <vijayendra.suman@xxxxxxxxxx> wrote:
> > >
> > >> Hello Mike,
> > >>
> > >> While Running pgbench tool with 5.4.17 kernel build
> > >>
> > >> Following performance degrade is found out
> > >>
> > >> buffer read/write metric : -17.2%
> > >> cache read/write metric : -18.7%
> > >> disk read/write metric : -19%
> > >>
> > >> buffer
> > >> number of transactions actually processed: 840972
> > >> latency average = 24.013 ms
> > >> tps = 4664.153934 (including connections establishing)
> > >> tps = 4664.421492 (excluding connections establishing)
> > >>
> > >> cache
> > >> number of transactions actually processed: 551345
> > >> latency average = 36.949 ms
> > >> tps = 3031.223905 (including connections establishing)
> > >> tps = 3031.402581 (excluding connections establishing)
> > >>
> > >> After revert of Commit
> > >> 2892100bc85ae446088cebe0c00ba9b194c0ac9d ( Revert "dm: always call
> > >> blk_queue_split() in dm_process_bio()")
> > >
> > >I assume 2892100bc85ae446088cebe0c00ba9b194c0ac9d is 5.4-stable's
> > >backport of upstream commit 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74 ?
> >
> > Yes
> >
> > >> Performance is Counter measurement
> > >>
> > >> buffer ->
> > >> number of transactions actually processed: 1135735
> > >> latency average = 17.799 ms
> > >> tps = 6292.586749 (including connections establishing)
> > >> tps = 6292.875089 (excluding connections establishing)
> > >>
> > >> cache ->
> > >> number of transactions actually processed: 648177
> > >> latency average = 31.217 ms
> > >> tps = 3587.755975 (including connections establishing)
> > >> tps = 3587.966359 (excluding connections establishing)
> > >>
> > >> Following is your commit
> > >>
> > >> diff --git a/drivers/md/dm.c b/drivers/md/dm.c
> > >> index cf71a2277d60..1e6e0c970e19 100644
> > >> --- a/drivers/md/dm.c
> > >> +++ b/drivers/md/dm.c
> > >> @@ -1760,8 +1760,9 @@ static blk_qc_t dm_process_bio(struct mapped_device
> > >> *md,
> > >> * won't be imposed.
> > >> */
> > >> if (current->bio_list) {
> > >> - blk_queue_split(md->queue, &bio);
> > >> - if (!is_abnormal_io(bio))
> > >> + if (is_abnormal_io(bio))
> > >> + blk_queue_split(md->queue, &bio);
> > >> + else
> > >> dm_queue_split(md, ti, &bio);
> > >> }
> > >>
> > >> Could you have a look if it is safe to revert this commit.
> > >No, it really isn't a good idea given what was documented in the commit
> > >header for commit 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74 -- the
> > >excessive splitting is not conducive to performance either.
> > >
> > >So I think we need to identify _why_ reverting this commit is causing
> > >such a performance improvement. Why is calling blk_queue_split() before
> > >dm_queue_split() benefiting your pgbench workload?
> >
> > Let me know if you want to check some patch.
>
> Hi,
>
> Could you please test this branch?:
> https://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm.git/log/?h=dm-5.10
> (or apply at least the first 4 patches, commit 63f85d97be69^..b6a80963621fa)
>
> So far I've done various DM regression testing. But I haven't tested
> with pgbench or with the misaaligned IO scenario documented in the
> header for commit 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74. But I'll
> test that scenario tomorrow.
Training DM core to set chunk_sectors and always use blk_queue_split
resolves the inefficient splitting documented in the header for
commit 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74.
xfs_io -d -c 'pread -b 2m 224s 4072s' /dev/mapper/stripe_dev
before, so with commit 120c9257f5f19e5d1e87efcbb5531b7cd81b7d74:
253,2 5 1 0.000000000 4382 Q R 224 + 2064 [xfs_io]
253,2 5 2 0.000003414 4382 X R 224 / 256 [xfs_io]
253,2 5 3 0.000017838 4382 X R 256 / 512 [xfs_io]
253,2 5 4 0.000019852 4382 X R 512 / 768 [xfs_io]
253,2 5 5 0.000031316 4382 X R 768 / 1024 [xfs_io]
253,2 5 6 0.000034333 4382 X R 1024 / 1280 [xfs_io]
253,2 5 7 0.000037684 4382 X R 1280 / 1536 [xfs_io]
253,2 5 8 0.000041011 4382 X R 1536 / 1792 [xfs_io]
253,2 5 9 0.000043962 4382 X R 1792 / 2048 [xfs_io]
253,2 5 10 0.000074765 4382 Q R 2288 + 2008 [xfs_io]
253,2 5 11 0.000075020 4382 X R 2288 / 2304 [xfs_io]
253,2 5 12 0.000077009 4382 X R 2304 / 2560 [xfs_io]
253,2 5 13 0.000080509 4382 X R 2560 / 2816 [xfs_io]
253,2 5 14 0.000084182 4382 X R 2816 / 3072 [xfs_io]
253,2 5 15 0.000087274 4382 X R 3072 / 3328 [xfs_io]
253,2 5 16 0.000090342 4382 X R 3328 / 3584 [xfs_io]
253,2 5 17 0.000095348 4382 X R 3584 / 3840 [xfs_io]
253,2 5 18 0.000097776 4382 X R 3840 / 4096 [xfs_io]
after, so with 'dm-5.10' branch refernced above, meaning dm_process_bio
w/ unconditional blk_queue_split (w/ chunk_sectors):
253,2 17 1 0.000000000 2176 Q R 224 + 2280 [xfs_io]
253,2 17 2 0.000001978 2176 X R 224 / 256 [xfs_io]
253,2 17 3 0.000017882 2176 X R 256 / 512 [xfs_io]
253,2 17 4 0.000020406 2176 X R 512 / 768 [xfs_io]
253,2 17 5 0.000031298 2176 X R 768 / 1024 [xfs_io]
253,2 17 6 0.000034654 2176 X R 1024 / 1280 [xfs_io]
253,2 17 7 0.000038474 2176 X R 1280 / 1536 [xfs_io]
253,2 17 8 0.000042299 2176 X R 1536 / 1792 [xfs_io]
253,2 17 9 0.000054088 2176 X R 1792 / 2048 [xfs_io]
253,2 17 10 0.000057884 2176 X R 2048 / 2304 [xfs_io]
253,2 17 11 0.000081358 2176 Q R 2504 + 1792 [xfs_io]
253,2 17 12 0.000081778 2176 X R 2504 / 2560 [xfs_io]
253,2 17 13 0.000083496 2176 X R 2560 / 2816 [xfs_io]
253,2 17 14 0.000085301 2176 X R 2816 / 3072 [xfs_io]
253,2 17 15 0.000092374 2176 X R 3072 / 3328 [xfs_io]
253,2 17 16 0.000094774 2176 X R 3328 / 3584 [xfs_io]
253,2 17 17 0.000097977 2176 X R 3584 / 3840 [xfs_io]
253,2 17 18 0.000100094 2176 X R 3840 / 4096 [xfs_io]
> Any chance you could provide some hints on how you're running pgbench
> just so I can try to test/reproduce/verify locally?
I'm going to defer to you on pgbench testing.
What is your underlying storage?
Could it be that DM using unconditional blk_queue_split() is helping
your pgbench workload because it splits IO more (so smaller IO, lower
latency per IO)?
Do you have comparison blktrace data?
Mike
