On Sun, 21 Jan 2007 16:04:44 -0500
"Perry E. Metzger" <perry@piermont.com> wrote:

> 
> "Steven M. Bellovin" <smb@cs.columbia.edu> writes:
> > I recently upgraded my kernel to one from January 8 (and then to one
> > from today).  Before that, I'd been running a mid-December -current.
> >
> > Interactive performance on my Thinkpad T42 has become completely
> > unacceptable.  Something else -- I have no idea what -- occasionally
> > grabs the CPU.  Moving the mouse shows jerky cursor motion, I get
> > doubled input characters, etc.  ps and top show that the machine is
> > mostly idle, though xosview sometimes shows brief spurts of 100%
> > busy. It isn't swapping (it's got 2G of RAM), the interrupt load
> > doesn't seem high (about 1.4-1.6K/sec, with HZ set to 1000), etc.  
> >
> > Is anyone else seeing this?  Anything I should do to tune it?  Might
> > the new scheduler code be part of it?
> 
> I can explain exactly what is going on but not why.
> 
> This sort of thing was very routine for me before before BUFQ_READPRIO
> (which was once called NEW_BUFQ_STRATEGY). Once I added that to my
> kernel years ago, the problem went away. I suspect you have that in
> your kernel.
> 
> As of Jan, something seems to have broken BUFQ_READPRIO -- unclear who
> did it or why, but it is very clear to me that it got broken. The
> old symptoms of lots of disk i/o locking out userland processes trying
> to read something have returned.
> 
I do indeed have BUFQ_READPRIO defined, but I'm far from convinced that
that's the problem -- this is happening without any disk-intensive
activity going on in the background.  Sure, there's always something
some daemon is doing, but I've tried shutting down other things.



		--Steve Bellovin, http://www.cs.columbia.edu/~smb