On Fri, 3 Mar 2000, Rik van Riel wrote:
> - the reports you saw here the last time contained error
> messages from kernels _without_ the "smart" killing logic
> - if you show me where the oom killing breaks for you, I'll
> gladly fix it
>
well, the reason i posted was because a machine running 2.2.14 that i'd
been logged into via XDM /suddenly/ started thrashing, intensive,
continious i/o, and xterms on that machine became very
unresponsive within a matter of 30 seconds. The machine was totally dead
about 5 min later, with lot's of killing xyz messages, where xyz were
things like fetchmail, crond, sendmail, etc... but not any of the X11
memory hogs (netscape)[1] that were actually causing the problem.
so i wasn't impressed.. however alan reckons 2.2.15 is improved, so i'll
wait and give that a try.
i still have a bad feeling about this killing thing though. I understand
it's half a dozen of one (harder VM, but possibly wasteful of RAM) and 6
of the other (less restrictive VM, more efficient RAM usage, but walking
on a tightrope). But i still think OOM killing is a kludge.
What'd be nice (perhaps for 2.5) would be perhaps VM policy classes. Eg
so that i could say "let processes of class A be allocated memory using
the standard 'lazy' VM model, and let the VM allocate memory to
processes in class B in a 'hard' way."
Processes in class A would ones whose memory needs/behaviour i
know/trust will not cause problems, eg most of my system daemons. Class
B processes would be those fat bloated hogs like netscape. How to define
certain processes being in one class, and others in another is a problem
though. And i gather from alan that 'hard' VMA is difficult to implement
correctly.
still, in a perfect world.... :)
> regards,
>
> Rik
> --
regards,
Paul Jakma.
paul@clubi.ie
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Tue Mar 07 2000 - 21:00:15 EST