arla-related hangs/pauses

[Nickolai Zeldovich]

I've been seeing occasional hangs in accesses to files/directories in
AFS lately on my FreeBSD 4.6.2 machine running arla 0.35.10pre4.  Any
process that tries to access AFS hangs for a long period of time, from
10-15 seconds to a few minutes.  Then all the processes un-freeze and
everything returns to normal for a while.  I suspect this is related
to the fact that Google is indexing my AFS cell through this machine,
but I hoped arla could handle this..

At first I tried increasing --workers from 16 to 64 (I got syslogs
about running out of workers), but it didn't help, and I don't get
such syslogs anymore.  I'm getting different messages now (see below)
but I suspect they aren't actually relevant (AFAICT they're because
something tried to fetch a file larger than the cache).

  Oct  8 07:34:36 orbit.zepa.net arla[81362]: Out of space since there are outstanding requests (335028224 needed, 0 outstanding, 314572800 highbytes
  Oct  8 09:40:48 orbit.zepa.net arla[81362]: Out of space, couldn't get needed bytes after cleaner (7702809 bytes missing, 25520409 used, 314572800 highbytes)
  Oct  9 04:06:03 orbit.zepa.net arla[81362]: Out of space, couldn't get needed bytes after cleaner (7458735 bytes missing, 17412015 used, 314572800 highbytes)
  Oct  9 05:07:13 orbit.zepa.net arla[81362]: Out of space since there are outstanding requests (324804608 needed, 0 outstanding, 314572800 highbytes

Any thoughts on what might cause such lock-up behavior, and if there
are any good ways to work around it (aside from "don't access AFS
so much")?

-- kolya