[lkml]   [2010]   [Feb]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [patch 4/7 -mm] oom: badness heuristic rewrite
On Wed, 10 Feb 2010, Rik van Riel wrote:

> > OOM_ADJUST_MIN and OOM_ADJUST_MAX have been exported to userspace since
> > 2006 via include/linux/oom.h. This alters their values from -16 to -1000
> > and from +15 to +1000, respectively.
> That seems like a bad idea. Google may have the luxury of
> being able to recompile all its in-house applications, but
> this will not be true for many other users of /proc/<pid>/oom_adj

Changing any value that may have a tendency to be hardcoded elsewhere is
always controversial, but I think the nature of /proc/pid/oom_adj allows
us to do so for two specific reasons:

- hardcoded values tend not the fall within a range, they tend to either
always prefer a certain task for oom kill first or disable oom killing
entirely. The current implementation uses this as a bitshift on a
seemingly unpredictable and unscientific heuristic that is very
difficult to predict at runtime. This means that fewer and fewer
applications would hardcode a value of '8', for example, because its
semantics depends entirely on RAM capacity of the system to begin with
since badness() scores are only useful when used in comparison with
other tasks.

- the badness() heuristic is radically changed from what it is currently
so this gives applications that hardcoded /proc/pid/oom_adj values into
their software a reason to notice the change and adjust to the new
semantics of the badness score. Using /proc/pid/oom_adj as a bitshift
has no real application to any sane heuristic that represents scores in
units of meaning, so users should end up with a net benefit of the
change by being able to better tune the oom killing behavior with a
much more powerful and easier to understand heuristic that requires
them to recalculate exactly what oom_adj should be for any given
application in terms of real units and business goals.

As mentioned in the changelog, we've exported these minimum and maximum
values via a kernel header file since at least 2006. At what point do we
assume they are going to be used and not hardcoded into applications?
That was certainly the intention when making them user visible.

> > +/*
> > + * Tasks that fork a very large number of children with seperate address
> > spaces
> > + * may be the result of a bug, user error, or a malicious application. The
> > oom
> > + * killer assesses a penalty equaling
> It could also be the result of the system getting many client
> connections - think of overloaded mail, web or database servers.

True, that's a great example of why child tasks should be sacrificed for
the parent: if the oom killer is being called then we are truly overloaded
and there's no shame in killing excessive client connections to recover,
otherwise we might find the entire server becoming unresponsive. The user
can easily tune to /proc/sys/vm/oom_forkbomb_thres to define what
"excessive" is to assess the penalty, if any. I'll add that to the
comment if we require a second revision.

Thanks for your speedy review of this patchset so far, Rik!

 \ /
  Last update: 2010-02-11 10:17    [W:0.101 / U:1.148 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site