> Inheriting points from the parent process doesn't really work very well.
> You could make an extremely small program, exec() it (which would
> presumably zero the score, or would you want to keep the score
> >from the large shell?),
Yes, since even a large shell should - when running for some time - have a
low score. And after all, if you *use* a large shell (emacs? ;-), it's
entirely your fault that the memory is so tight :-).
> Evanescating scores by a certain factor per second just means that a
> malicious hog has to be careful to wait for a certain time before
> starting.
Hm, one point is: if a hog allocates memory slowly, you have time to
react. After all, there are a lot of non-malicious programs that allocate
memory slowly over time, and you are really pissed if they get killed.
> Also it's not good if two processes allocate the same amount of memory
> shortly after one another, but only the second gets malus points since
> swapping only started when the second one allocated.
This is just a sort of conservative capitalistic/aristocratic algorithm.
If all cheap resources are occupied, allocating new resources becomes more
expensive. The idea is to keep the setting in a sane state by killing
newcomers. People don't want to get their two-week simulation killed,
therefore Rik's algorithm takes total cpu usage (processes with more
computing power invested are worth more) and runtime into account.
I think Rik's algorithm is good enough and has the enormous advantage of
being cost-free in normal operation (*and* it has been tested). I'll take
his weighting algorithm (size/(sqrt(cpu_time)*sqrt(sqrt(run_time)))), but
add my warning shot, and convert it to a more recent kernel (2.3.10). Too
bad that we don't have a free signal for SIGMEM at hand ;-).
Bernd Paysan
"If you want it done right, you have to do it yourself"
http://www.jwdt.com/~paysan/
-
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/