[OpenAFS] Salvager did not run automatically on solaris 9, 1.4.1-rc10

Renata Maria Dart renata@slac.stanford.edu
Fri, 14 Apr 2006 11:53:27 -0700 (PDT)


On Thu, 13 Apr 2006, Jeffrey Hutzelman wrote:

>Under the circumstances you describe, yes, this is normal.
>
>However, you switched to namei, which doesn't have that feature (and can't,
>since it doesn't use a modified fsck).  So on your first start, the
>bosserver forced a salvage, but there weren't any partitions to salvage, so
>nothing interesting happened.  Then the fileserver started up, and someone
>noticed there were no partitions, and rebooted.  That involved a clean
>shutdown of the fileserver, which meant there was no forced salvage on the
>next boot.
>
>
>I have to admit I'm a little curious why you switched from inode to namei
>on a Solaris server...
>


Thanks, I figured that was what was going on.  As far as namei vs
inode, for the person who just wrote in and was worried about the fsck
issue, we have been running inode here for years with no
problems....there has never been a mistake with fsck.  I decided to
investigate namei for the reasons that follow and so I have deployed
it on a couple of our servers, but the majority continue to run inode
and we may never switch to namei.

A few years ago we deployed a Sun tool called liveupgrade to
make nightly copies of the boot drive to an alternate on many of our
critical systems, for quick recovery in case of bootdrive failure.
This has worked nicely for us, but Sun has changed the way liveupgrade
works and now it has a problem with any fstypes that it does not
bless, such as afs.  Even though I don't use liveupgrade to copy any
afs filesystems, it objects to their existence in the vfstab.

In part because of the above experience, it seems sensible to stick as
much as possible with the vendor supported filesytem.  And with zfs
looming on the horizon, since that is supposed to be the filesystem of
one's dreams, I would like to move to it when it becomes available and
thought I would be forced to namei at that point anyway, although now
it sounds like there is a chance you may get inode working with that
too.

Of the systems that I have converted to namei, one is a lightly loaded
V240 with a T4 attached and the other is a 280R with a T3.  While
performance has been the reason why we have stuck with inode, as
systems and RAID get faster I thought the performace disparity between
the two should decrease.  A couple of years ago Alf Wachsmann ran some
benchmarks to compare inode and namei which according to his results
found inode to be 20% better than namei on logging UFS, for some
operations.  But his tests were done on a netra T1 with an A1000.  (I
should try and wheedle him into doing a similar test with the faster
hardware that we have now.)  So I thought surely our current systems
should perform better and namei will prove to be closer to inode in
performance.

With namei you get to use Solaris logging which I was hoping would cut
down on the length of time it takes to salvage.

Do you have any recommendations on namei vs inode on solaris?  Am I
expecting too much of our faster systems as far as the performance hit
we might see by switching to namei (I couldn't find a benchmark out
there for this hardware).  Is there anything other than performance
which would be lost and should be considered when moving from inode to
namei?

Thanks,

Renata