Re: [HTCondor-users] Out of memory killer & cgroups

[Greg Thain <gthain@xxxxxxxxxxx>]

On 10/02/2014 09:19 AM, Rich Pieri wrote:
> The Linux kernel OOM killer is independent of cgroups. The only 
> interaction with cgroups is that a cgroup memory limit may cause the 
> OOM killer to activate sooner than it would without a constrained 
> memory limit. SIGKILL (kill -9) is immediate and it cannot be trapped 
> or ignored. The killed process does not have a chance to write out any 
> logs or otherwise clean up after itself so there's nothing that it can 
> do to let users know why it was killed. What the parent does is up to 
> the parent, although right off it has no way to distinguish between a 
> KILL signal sent by the kernel and a KILL signal sent by a user. So, 
> on the face of it, the behavior that you are seeing is something that 
> I would expect to see. Whether or not it's the intended behavior is 
> something that I will leave to the Condor devs to address. 

Note this isn't entirely correct.  A process can register to have the 
cgroup memory controller notify it when the per-cgroup OOM fires.  This 
is what the condor_starter does, so that it can differentiate between 
the OOM killer firing, and some other reason the job was kill'ed -9.

-Greg