OSEC

Neohapsis is currently accepting applications for employment. For more information, please visit our website www.neohapsis.com or email hr@neohapsis.com
 
Subject: Re: VERY BAD performances (help !)
From: Bennett Todd (betrahul.net)
Date: Wed Apr 12 2000 - 09:02:22 CDT

2000-04-11-22:22:47 Wietse Venema:
> > > As long as the file system correctly reports out-of-space
> > > conditions, otherwise it will be just like the optimistic
> > > memory allocator in AIX that would always succeed.
> >
> > How would that system recover from out of {memory,disk space}
> > conditions? Crash hard with SIGSEGV, SIGABRT or something? Even
> > send SIGKILL for resource exhaustion?
>
> malloc() would always succeed, and the process would experience
> SIGSEGV a while later. Makes it a real challenge to write robust
> applications - it was impossible to distinguish between a genuine
> programming bug and an elementary out of memory condition.

Very interesting. This sounds similar, but not identical, to
a big battle that raged over VM systems about a decade ago or
thereabouts. Way back when, BSD had a memory allocation strategy
that treated swap space as VM, and real memory was just a cache onto
the swap. All allocations were committed out of swap, only. BSD
robustly reported allocation failures at allocation request time,
and no bad craziness happened later. Meanwhile System V had a memory
allocation system that juggled as best it could, and presented VM
is the sum of available swap and real. As memory got cheaper, going
with swap-less systems became more and more desireable, so the
System V strategy came into favour.

The big complaint about at the time was that unlike the traditional
BSD allocator, the System V approach could, in an unlikely sequence
of events, get itself deadlocked where its only way to proceed would
be to elect a victim and kill it off.

But the above tale is somewhat different from what you described;
the "System V" strategy _almost_ always correctly reported
allocation failures at request time, and in practice I never heard
of anyone catching it killing off a victim to recover from its
memory allocation shenanighans.

-Bennett

  • application/pgp-signature attachment: stored