On Sun, Sep 19, 1999 at 12:15:10PM -0500, Jesse Pollard wrote:
[snip]
> Only a single compair per resource should be needed. Now, if the system is
> oversubscribed, there may need to be more.. but only (political) management
> can force oversubscription. And if they do, then they deserve the system
> hang/crash that may occur. BTW, that really does happen. We had a Cray
I've got your idea. But _all_ real systems are oversubscribed.
There are 15 users on my working system. Each of them is allowed to spawn
256 processes. Each process is allowed to consume up to 128M memory.
The total amount of memory is 128M too. So my system is 384000%
oversubscribed. That's how I have always been managing my system and want to
manage it in the future. That's how 99.9% administrators do.
The proposed system of per-user resource accounting reduces the risk of crash
to an acceptable value. Nothing less is acceptable, nothing more is
necessary.
> system that had swap oversubscribed - everytime someone filled the actual swap
> space the system would hang. And it kept happening until the oversubscription
> was eliminated. This problem also occurs on SGI Origin systems that are
> oversubscribed, but there the processes get killed - even system processes
> like inetd, getty, init.. anything that attempts to use virtual memory.
>
> >So much limits is really an unnecessary overhead.
>
> It is "unnecessary" only if you do not have to justify the next upgrade
> of a 100 system beowulf cluster (with possibly 200 CPUs, and 100 GB of
> memory), or justify a larger allocation of an existing system.
>
> Unnecessary is relative to the size of the system.
For me 10% of performance degradation without a good reason is unacceptable
independently of the system size.
Regards
Andrey V.
Savochkin
-
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/