[ILUG] Out of Memory

[Paul Jakma]

On Mon, 6 Mar 2000, David Murphy wrote:

> By this logic, the fact that Win 3.1 program can crash the whole
> machine doesn't make it a bad OS - it's just a bug in the programmer.
> It's a fact of life that we expect OSes to survive bugs in programs
> run under them.
>

well, it's not a bug in linux per se, it's just linux sails a lot closer
to the wind viz VM allocation than other OS's. And therefore linux gets
caught out more often than other OS's, but it's not an OS bug, it's an
OS design compromise - solaris can OOM too.

also, the OOM killing in 2.2.15 should be a lot lot better, eg it won't
kill processes that hold iopl() like X. (but that's post OOM behaviour -
if you've seen solaris killing processes, then it was in the same boat
as linux).
  
> Is that a single-processor Alpha, by any chance? How about a 4-way
> one? An 8-way one? a 16-way one? (etc.).
>

1/2 cpu of course.. I don't think linux will ever scale to 32+ cpu's as
good as sol/t64/irix. Then again, you pay heavily for that scaleability
in performance lost on lower end systems..
 

> IMHO. Solaris for Merced/Itanium is the first non-SPARC version Sun
> are even pretending to take seriously.
> 

well linux romps way ahead of Sol on 1way uSparc too from what i've
read. (sol x86/t64 are the only ones i've experience of Vs linux).

> I believe that may well be the practical reality, but Linux is being
> sold as enterprise-ready - the caveat you cite would be appropriate
> for a research OS.
> 

but that caveat exists in all OS's that allow overcommit. And that
includes default installs of solaris and T64 and IRIX. 

Of course, if you wanted you could configure linux/intel with 4GB of
swap. That just about guarantees OOM will never happen. (of course t64
and probably solaris too need tonnes of swap aswell if you configure
them for restrictive memory mgt...)

-paul.