> No swap means that no charges are done by a group non-member. So the > race Johannes was describing shouldn't be the problem in your case. > > Out of curiosity, do you set any limit for kmem? > I did a quick search through the code, but I see nothing related to kernel memory: https://github.com/SchedMD/slurm/blob/master/src/plugins/task/cgroup/task_cgroup_memory.c >> memory.use_hierarchy support is enabled and memory is limited from the >> job directory. I have mounted the cgroup to /var/slurm/cgroup/memory >> in addition to the normal directory at /sys/fs/cgroup/memory. > > How exactly have you mounted it there? Slurm has an automount option, which takes care of this. I don't actually know, why I made this kind of setup. So I will probably revert back to /sys/fs/cgroup tomorrow. > > Btw. how reproducible is this? Do you think you could try to bisect > it down? Reducing bisection to mm/ and kernel/ diretories should be > sufficient I guess. The bug is quite reproducible here .. within a few minutes at most. Since we have diskless clients with nfsroot and aufs, bisectioning proved to be a bit difficult (means the kernel compiled and booted but aufs failed on /etc). But then I have tried bisectioning the whole kernel sources. Tomorrow, I will first try to test without CONFIG_MEMCG_SWAP_ENABLED and CONFIG_MEMCG_KMEM, and then give bisectioning only the two directories another try. Btw. I never saw this bug before 3.11, but it may well be, that because of some trivial code change, it became much more likely to trigger. -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
