On Tue, Oct 30, 2007 at 07:49:41PM +0800, Fengguang Wu wrote:
quoted text
> > > It could be triggered by the more aggressive writeback behavior - the
> > > new code will keep on retrying as long as there are dirty inodes pend=
ing.
quoted text
> > >=20
> > > Florin, would you try the attached patches against 2.6.24-git?
> > > They may generate big traffic of printk messages, but will help
> > > debug the problem.
> >=20
> > I have updated to v2.6.24-rc1-334-g82798a1.  After using my computer
> > for two hours, I left the computer idle overnight.  This morning,
> > pdflushd is again consuming 25% of a CPU.  I will try Fengguang's
> > patches today.
>=20
> Thank you.
>=20
> Be sure to stop klogd ;-)
> A `dmesg` will be sufficient for debugging.

I have added the patches and started a linux kernel compilation, and
something really interesting happens.  I run the build with the
equivalent of "make -j3" and in a separate console I am watching the
build with 'top'.  The build consumes 98% of both CPUs.  If I stop the
output in the build console with "Ctrl-S", one core goes to idle,
while the other is in 50% waiting, then goes to 75% waiting.  When I
resume the build with "Ctrl-Q", the build starts to use both CPUs at
98-99%.  The NFS4 use was minimal, as I did not login with Gnome, but
just logged on the console.  Also, the CPU that is in 75% waiting
state changes occasionally.  'Top' shows pdflush in D state, using
5-6% of CPU.

Cheers,
florin

--=20
Bruce Schneier expects the Spanish Inquisition.
      http://geekz.co.uk/schneierfacts/fact/163