Re: buffer/swapping in pre-patch-2.0.31-2 + the 17th July patch
Richard Waltham (dormouse@farsrobt.demon.co.uk)
Wed, 23 Jul 1997 20:24:57 +0100 (BST)
Dr. Werner Fink wrote:
>
> > >looks like I'm catching this one:
> >
> > >Jul 19 16:07:15 arak kernel: try_to_free_page: free pages 1008kB, async pages 0kB
> > >Jul 19 16:08:56 arak kernel: try_to_free_page: free pages 1020kB, async pages 12kB
> > >Jul 19 16:12:04 arak kernel: try_to_free_page: free pages 1012kB, async pages 4kB
>
> > Same here:
> >
> > Jul 21 07:57:25 smithers kernel: try_to_free_page: free pages 776kB, async pages 0kB
> > Jul 21 07:57:25 smithers kernel: Appletalk 0.17 for Linux NET3.035
> > Jul 21 09:15:08 smithers kernel: try_to_free_page: free pages 396kB, async pages 0kB
> > Jul 21 09:15:09 smithers kernel: try_to_free_page: free pages 384kB, async pages 0kB
> > Jul 21 09:15:29 smithers kernel: try_to_free_page: free pages 436kB, async pages 0kB
> > Jul 21 09:15:30 smithers kernel: try_to_free_page: free pages 448kB, async pages 0kB
>
>
> Hmmm ... interesting. Please try the appended patch. It contains a few
> changes:
>
> * run buffer wait queue only if recover_reusable_buffer_heads()
> is called in brw_page() due get_more_buffer_heads()
>
> * Set PAGE_AGE_VALUE to PAGE_INITIAL_AGE+PAGE_ADVANCE in
> linux/include/linux/pagemap.h ... should avoid over-aged
> cache pages
>
> Are there any differences in bonnie? Is the debugging printk in
> try_to_free_page() triggered?
>
> Werner
>
Hi Werner,
The printk hasn't appeared with this one or the latest July 22 patch.
The following comments on bonnie apply equally to 17, 21 and 22 July
patches.
bonnie looks pretty similar but FEELS a little sluggish especially once
multiple copies of bonnie are running or multiple copies of
writing/comparing largish files ~100MBytes across a pair of disks. I
haven't checked the results of the write/compares yet. I suppose this could
be due to profiling running.
writing is a little slower than plain 2.0.29, about 5% slower. Reading is
comparable with the better results although July 22 version may just be
marginally better than previous 2 patches.
Best writes speeds were with your 10 July patch 15-20% faster than at
present.
Worst was 8 July where sequential reads dropped to ~500KBs :(:(:(
If you or anyone else is interested in/could make use of strace summaries
or profiling info I have samples from 2.0.29 onwards although appart from a
couple of noticable exceptions they all look much the same to me.
Richard
--
Richard Waltham | Work: richard@digtalinterface.demon.co.uk
At home in | Home: dormouse@farsrobt.demon.co.uk
Southampton UK | Compuserve 100421.1276