Re: stress testing and loadavg

David S. Miller (davem@jenolan.rutgers.edu)
Thu, 7 Aug 1997 19:24:19 -0400


Date: Thu, 7 Aug 1997 17:15:19 +0200 (MET DST)
From: "Michael L. Galbraith" <mikeg@weiden.de>

Would 'interesting things' happen to be..
_setutent: Can't open utmp file: File table overflow
with top segfaulting and (sometimes) the make -j also segfaulting?
The xterm was permanently broken, outputing..
sh: command_substitute: cannot duplicate pipe as fd1: Bad file number
The same thing happened to a shell (make -j) in a normal vt, but after a short
while, it started working again. (I mean _after_ killing all fun_programs)

Yep I get this too. I know a lot of it has to do with the fork
bombers eating up the 2-page order pools for kernel stack/task_struct
to the point where memory becomes too wildly fragmented to satisfy any
further requests, this needs to be fixed for eventual 2.2.0.

The file table overflow is queer, I saw them frequently too. If I
killed everything off and waited a bit, that behavior went away. It
may be another case of allocation failures, but I never investigated
it deeply enough to find out whether this was truly the case.

Linus should run these tests a bit and have a look at what it does to
the kernel ;-)

Later,
David "Sparc" Miller
davem@caip.rutgers.edu