On Fri, May 13, 2011 at 4:38 AM, David Rientjes<rientjes@xxxxxxxxxx> wrote:On Thu, 12 May 2011, Minchan Kim wrote:
processes a 1% bonus for every 30% of memory they use as proposed
earlier.)
I didn't follow earlier your suggestion.
But it's not formal patch so I expect if you send formal patch to
merge, you would write down the rationale.
Yes, I'm sure we'll still have additional discussion when KOSAKI-san
replies to my review of his patchset, so this quick patch was written only
for CAI's testing at this point.
In reference to the above, I think that giving root processes a 3% bonus
at all times may be a bit aggressive. As mentioned before, I don't think
that all root processes using 4% of memory and the remainder of system
threads are using 1% should all be considered equal. At the same time, I
do not believe that two threads using 50% of memory should be considered
equal if one is root and one is not. So my idea was to discount 1% for
every 30% of memory that a root process uses rather than a strict 3%.
That change can be debated and I think we'll probably settle on something
more aggressive like 1% for every 10% of memory used since oom scores are
only useful in comparison to other oom scores: in the above scenario where
there are two threads, one by root and one not by root, using 50% of
memory each, I think it would be legitimate to give the root task a 5%
bonus so that it would only be selected if no other threads used more than
44% of memory (even though the root thread is truly using 50%).
This is a heuristic within the oom killer badness scoring that can always
be debated back and forth, but I think a 1% bonus for root processes for
every 10% of memory used is plausible.
Comments?
Yes. Tend to agree.
Apparently, absolute 3% bonus is a problem in CAI's case.
Your approach which makes bonus with function of rss is consistent
with current OOM heuristic.
So In consistency POV, I like it as it could help deterministic OOM policy.
About 30% or 10% things, I think it's hard to define a ideal magic
value for handling for whole workloads.
It would be very arguable. So we might need some standard method to
measure it/or redhat/suse peoples. Anyway, I don't want to argue it
until we get a number.