Re: [performance bug] kernel building regression on 64 LCPUs machine

["Alex,Shi" <alex.shi@intel.com>]

Sorry for forgetting Jan and tytso. added again. 

On Fri, 2011-01-21 at 15:47 +0800, Alex,Shi wrote:
> On Fri, 2011-01-21 at 15:23 +0800, Corrado Zoccolo wrote:
> > On Wed, Jan 19, 2011 at 2:55 AM, Alex,Shi <alex.shi@intel.com> wrote:
> > > Shaohua and I tested kernel building performance on latest kernel. and
> > > found it is drop about 15% on our 64 LCPUs NHM-EX machine on ext4 file
> > > system. We find this performance dropping is due to commit
> > > 749ef9f8423054e326f. If we revert this patch or just change the
> > > WRITE_SYNC back to WRITE in jbd2/commit.c file. the performance can be
> > > recovered.
> > >
> > > iostat report show with the commit, read request merge number increased
> > > and write request merge dropped. The total request size increased and
> > > queue length dropped. So we tested another patch: only change WRITE_SYNC
> > > to WRITE_SYNC_PLUG in jbd2/commit.c, but nothing effected.
> > >
> > > we didn't test deadline IO mode, just test cfq. seems insert write
> > > request into sync queue effect much read performance, but we don't know
> > > details. What's your comments of this?
> > >
> > > iostat of .37 kernel:
> > > rrqm/s   wrqm/s r/s     w/s     rMB/s  wMB/s avgrq-sz avgqu-sz   await  svctm  %util
> > > 22.5772 96.46 92.3742 14.747  1.0048  0.439474 34.8557 0.18078 3.8076 0.30447 2.94302
> > > iostat of commit reverted .37:
> > > 26.6223 80.21 107.875 6.03538 1.51415 0 41.3275 0.153385 3.80569 0.377231 3.22323
> > 
> > From these numbers, it seems to me that with the patch reverted, the
> > write bandwidth is really low, and probably you are keeping most
> > written files in the buffer cache during the whole compile, while the
> > non-reverted kernel is making progress in writing out the files. So
> > the 'improved' read bandwidth is due to unfairness w.r.t. writes.
> > Does the result change if you add a final sync and time that as well,
> > in order to see the full time to make it on disk?
> 
> Agree with your guess, but kbuild is such kind of benchmark, we can not
> change its behavior. :( 
> 
> 
> > 
> > I think that in a more extreme test where you end up filling all the
> > buffer cache with written data, you will see much longer stalls with
> > the revert than without.
> 
> Have to do this? and if so, it is not kbuild. :) 
> 
> BTW, the Jan's patch has a little improvement on kbuild. In many time
> testing, it seems about 3% improving. 
> 
> The average iostat output of Jan's patch: 
>  rrqm/s   wrqm/s r/s     w/s     rMB/s  wMB/s avgrq-sz avgqu-sz   await  svctm  %util
> 23.419726 87.450685 96.164521 6.748493 1.046438 0.370137 45.182192 0.200685 6.848767 0.394110 3.072192
> 
> > 
> > Thanks,
> > Corrado
> > 
> > >
> > > vmstat report show, read bandwidth dropping:
> > > vmstat of .37:
> > >  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa st
> > > 3.4 52.6 0.0 64303937.0 16466.7 121544.5 0.0 0.0 2102.7 1914.6 7414.1 3185.7 2.0 1.0 80.3 16.7 0.0
> > > vmstat of revert all from .37
> > > 2.2 35.8 0.0 64306767.4 17265.6 126101.2 0.0 0.0 2415.8 1619.1 8532.2 3556.2 2.5 1.1 83.0 13.3 0.0
> > >
> > > Regards
> > > Alex
> > >
> > > ===
> > > diff --git a/fs/jbd/commit.c b/fs/jbd/commit.c
> > > index 34a4861..27ac2f3 100644
> > > --- a/fs/jbd/commit.c
> > > +++ b/fs/jbd/commit.c
> > > @@ -294,7 +294,7 @@ void journal_commit_transaction(journal_t *journal)
> > >        int first_tag = 0;
> > >        int tag_flag;
> > >        int i;
> > > -       int write_op = WRITE_SYNC;
> > > +       int write_op = WRITE;
> > >
> > >        /*
> > >         * First job: lock down the current transaction and wait for
> > > diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c
> > > index f3ad159..69ff08e 100644
> > > --- a/fs/jbd2/commit.c
> > > +++ b/fs/jbd2/commit.c
> > > @@ -329,7 +329,7 @@ void jbd2_journal_commit_transaction(journal_t *journal)
> > >        int tag_bytes = journal_tag_bytes(journal);
> > >        struct buffer_head *cbh = NULL; /* For transactional checksums */
> > >        __u32 crc32_sum = ~0;
> > > -       int write_op = WRITE_SYNC;
> > > +       int write_op = WRITE;
> > >
> > >        /*
> > >         * First job: lock down the current transaction and wait for
> > >
> > >
> > >
> > 
> > 
> > 
>