On Mon, Aug 16, 2010 at 09:20:05AM -0400, Jeff Moyer wrote:
quoted text
> Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> writes:
> 
> > Hi,
> >
> > On Thu, Aug 12, 2010 at 09:42:21AM -0400, Jeff Moyer wrote:
> >> Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> writes:
> >> 
> >> > Basically it is user's responsibility to take care of race condition
> >> > related to direct I/O, but some events which are out of user's control
> >> > (such as memory failure) can happen at any time. So we need to lock and
> >> > set/clear PG_writeback flags in dierct I/O code to protect from data loss.
> >> 
> >> Did you do any performance testing of this?  If not, please do and
> >> report back.  I'm betting users won't be pleased with the results.
> >
> > Here is the result of my direct I/O benchmarck, which mesures the time
> > it takes to do direct I/O for 20000 pages on 2MB buffer for four types
> > of I/O. Each I/O is issued for one page unit and each number below is
> > the average of 25 runs.
> >
> >                                   with patchset          2.6.35-rc3
> >    Buffer      I/O type        average(s)  STD(s)   average(s)  STD(s)   diff(s)
> >   hugepage   Sequential Read      3.87      0.16       3.88      0.20    -0.01
> >              Sequential Write     7.69      0.43       7.69      0.43     0.00
> >              Random Read          5.93      1.58       6.49      1.45    -0.55
> >              Random Write        13.50      0.28      13.41      0.30     0.09
> >   anonymous  Sequential Read      3.88      0.21       3.89      0.23    -0.01
> >              Sequential Write     7.86      0.39       7.80      0.34     0.05
> >              Random Read          7.67      1.60       6.86      1.27     0.80
> >              Random Write        13.50      0.25      13.52      0.31    -0.01
> >
> > From this result, although fluctuation is relatively large for random read,
> > differences between vanilla kernel and patched one are within the deviations and
> > it seems that adding direct I/O lock makes little or no impact on performance.
> 
> First, thanks for doing the testing!
> 
> > And I know the workload of this benchmark can be too simple, so please
> > let me know if you think we have another workload to be looked into.
> 
> Well, as distasteful as this sounds, I think a benchmark that does I/O
> to partial pages would show the problem best.  And yes, this does happen
> in the real world.  ;-)  So, sequential 512 byte or 1k or 2k I/Os, or
> just misalign larger I/Os so that two sequential I/Os will hit the same
> page.
> 
> I believe you can use fio to generate such a workload;  see iomem_align
> in the man page.  Something like the below *might* work.  If not, then
> simply changing the bs=4k to bs=2k and getting rid of iomem_align should
> show the problem.

Thank you for information.

I measured direct I/O performance with small blocksize or misaligned setup.
The result is shown here:

                                 average bandwidth
                        with patchset       2.6.35-rc3     diff
    bs=512                1,412KB/s          1,789KB/s    -26.6%
    bs=1k                 2,973KB/s          3,440KB/s    -13.6%
    bs=2k                 6,895KB/s          6,519KB/s     +5.7%
    bs=4k                13,357KB/s         13,264KB/s     +0.7%
    bs=4k misalign=2k    10,706KB/s         13,132KB/s    -18.5%

As you guessed, the performance obviously degrades when blocksize is small
and when I/O is misaligned.

BTW, from the discussion with Christoph I noticed my misunderstanding
about the necessity of additional page locking. It would seem that
without page locking there is no danger of racing between direct I/O and
page migration. So I retract this additional locking patch-set.

Thanks,
Naoya Horiguchi
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/