ZFS & Bittorent -> Hang?

[James Snyder]

Thanks for the followup.  I have not yet gotten a reliable test case
to reproduce the problem. I've done a number of tests with the zil
and/or prefetch on with no hangs.  I will be collecting some more data
later this week.

If anyone knows a source of consistently slow but large torrents (I
suppose I could artificially limit bandwidth, or connection states at
my router which is running pfSense), that might help for testing.  The
process that triggered things before was about a gigabyte or two but
took around 12 hours to complete.

Here's the overall group of variables I'm experimenting with.

Stock Kernel vs Recompiled Kerel w/ ULE (stock sources otherwise)
ZIL on and off
prefetch on and off

Should I add or remove anything?  I have no idea if ULE may or may not
play a role here, but my original failing condition had the zil off,
prefetch on, ule for the scheduler.

I also had:
vm.kmem_size_max="1073741824" (loader.conf)
vm.kmem_size="1073741824" (loader.conf)

Any recommendations on what to leave running to record what zfs is
getting hung on, beside watching states?  Since I can fire up things
prior to the hang, and they'll keep running if disk isn't hit, I could
leave some diagnostics running to display what's blowing up.

Thanks!

On Tue, Apr 15, 2008 at 4:44 AM, Claus Guttesen <kometen at gmail.com> wrote:
> >  > http://wiki.freebsd.org/ZFSKnownProblems
>  >  >
>  >  > This looks like #1.
>  >  >
>  >
>  >  Hmm.. I don't think there's a large amount of transfer between UFS & ZFS,
>  >  unless the client is using /tmp a lot, it should all be on ZFS.
>  >
>  >  I noted #4 as well, and therefore tried disabling prefetch.  I can't seem to
>  >  get it to hang now.  I queued up a bunch of different torrents (full freebsd
>  >  7 amd64 & i386, some other random things), and they all completed without
>  >  leading to me being locked out or any processes waiting on zfs.
>  >
>  >  I'll try some testing this weekend to see if I can reproduce the lock-up
>  >  again by re-enabling prefetch.  Perhaps we can confirm that issue?  Should I
>  >  bother with trying to run CURRENT or should any testing I do be done with
>  >  STABLE.  I don't see any indication that there might be experimental patches
>  >  for dealing with this or related issues.
>
>  Were you able to reproduce the lock-up by re-enabling prefetch?
>
>  --
>  regards
>  Claus
>
>  When lenity and cruelty play for a kingdom,
>  the gentlest gamester is the soonest winner.
>
>  Shakespeare
>



-- 
James Snyder
Biomedical Engineering
Northwestern University
jbsnyder at gmail.com