Thanks for pointing this out. The patch is at the link below and will be in Slurm version 2.4.5 and 2.5.0. It should also go into earlier versions of Slurm cleanly if desired.
https://github.com/SchedMD/slurm/commit/8ae5e73e26e3e37deec85c644afcf30fb0bb1e7b Quoting Michel Bourget <[email protected]>: > > Hi, > > since oom_adj is deprecated since 2.6.36 kernel ( see > http://code.google.com/p/chromium/issues/detail?id=65009 among others ) > and semantic has changed, ie.: > > /proc/pid/oom_adj range: -17..15 > /proc/pid/oom_score_adj range: -1000..1000 > > This means writing -17 on /proc/pid/oom_score_adj does not offer oom > protection > at all. It should be -1000. > > I guess it's a matter of documentation and adjustments: > - SLURMD_OOM_ADJ and SLURMSTEPD_OOM_ADJ envars > - adjust set_oom_adj() to do "The Right Thing(Tm)" depending > on oom_score_adj presence or not. > - doc/html/faq.shtml adjustments ... > - src/plugins/task/cgroup/task_cgroup_memory.c adjustments: > it use -17 but that may end up in oom_score_adj, hence not > oom protecting ... > > And, yes, "I hate when they do that(Tm)" :) > > FYI, I checked slurm 2.3.4 source. > Maybe, it is already handled in recent slurm versions ? > > > A+ > > -- > > ----------------------------------------------------------- > Michel Bourget - SGI - Linux Software Engineering > "Past BIOS POST, everything else is extra" (travis) > ----------------------------------------------------------- >
