Subject: Re: NFS swap problems?
To: None <tech-kern@NetBSD.ORG>
From: der Mouse <mouse@Collatz.McRCIM.McGill.EDU>
List: tech-kern
Date: 08/22/1995 20:46:53
> [ NOTE: this may sound sun3-related, but I'm almost willing to bet
>   it's something in MI code causing this. ]

This, I can't say much about, but what little I can say indicates that
you're right.  See below.

> At work, I have a Sun 3/260 running diskless (because I haven't
> installed the latest-greatest sun3 boot code yet) that, under heavy
> load (well, building the tree), hangs solid from time to time.  I
> noticed this at home on my 3/50 before I installed a disk on it.

> JT mentioned that this has happened to him, as well.  It seems to
> only happen on diskless machines with small memories while doing
> something.

My experience is a little different.  I have a -3/150 (basically, a
-3/160 with a smaller cardcage), which I was running diskless off my
NeXT.  When building the tree, it would lose randomly.  I mean things
like make mysteriously complaining it was unable to allocate memory.
Like nm coredumping in the "ar `lorder *.o | tsort`" phase of library
building.  Like inexplicable and unrepeatable errors from cc.

I never saw it hang.  I just saw the build croak.  I eventually gave up
on it, and am holding off on building anything there until I get some
disk on that box (which I'm hoping to get around to soon).

Interestingly, quite a while ago I did a full build on a -3/260,
diskless off the same NeXT, with no problems.  It ran for days on end,
plugging doggedly away.  But on the -3/150, with a modern kernel, I see
problems.  I don't know whether it's the cpu or the updated kernel that
makes the difference; I do know that the successful build was running
under a kernel without the cache-enabling code for 260s (and also
before the zs driver fixes, which is why it was on the 260 rather than
the 150).

Interestingly, remember that code I mentioned that provided unionfs-ish
semantics via an NFS mount?  When I tried doing a full build on top of
that, on the SPARC, I saw similar problems.  All the disk was in fact
local, but /usr/src was, as far as the kernel was concerned, an NFS
mount.  (The kernel was even more current than the sun3 kernel I was
using on the 150.)

					der Mouse