On Sun, 5 Jul 2015, Dave Love wrote:
...
I don't understand. If you drop the rlimit, the only control of memory
is qmaster killing when the measured usage exceeds the h_vmem. In
particular, the application can't fail gracefully if it tries to use too
much memory, and it can cause OOM by allocating fast enough, possibly
getting another job killed that is within limits. (One of the rubbish
things about cgroups is that you can't fail gracefully on memory
allocation. Another is the performance penalty, though I haven't
measured that for relevant applications.)
Hi Dave,
I agree, which is why my original memory cgroup patchset separated out
control of RAM from virtual address space usage with a new resource: that
way people who care about such things could revert back to the old
behaviour. I don't know if that made sense for the other platforms
supported by soge - I probably went a bit crazy (I think I provided h_as
for rlimit control and renamed h_vmem to h_mem for the memory cgroup).
But at present the, sometimes wildly inaccurate, per-process control
provided by rlimit tends to trump the wonderfully accurate per-job control
provided by the execd polling PSS. Which makes the PSS polling pointless
just when you want it the most, in many common cases.
It seems to me that those who are conservative about at what point exactly
a job is killed have their needs met, but those who want to efficiently
use the available hardware with applications that have an abnormally large
address space compared to RAM utilisation (mmap, java, etc.) still do not.
What is the "realistic measurement by the kernel"?
I was referring to the PSS work you did.
...
The whole cgroups thing is a horrible moving target. I think it was a
mistake to bake in support, rather than providing appropriate hooks like
SLURM's, rather than the re-write I'd started on. That's not likely to
get done, though, unless someone else takes over, especially as I
couldn't test it properly. People should probably move to SLURM for
better resource management.
I think a lot of people have a love/hate relationship with the kernel's
cgroup feature :)
You have my gratitude for everything you've done with soge - I for one am
certainly glad you put in cpuset support.
Mark
_______________________________________________
SGE-discuss mailing list
[email protected]
https://arc.liv.ac.uk/mailman/listinfo/sge-discuss
