On Tue, May 22, 2012 at 2:02 PM, William Hay <[email protected]> wrote:
> Well currently we have our prolog chgrp the appropriate number of
> /dev/nvidia? files to the group associated with the job thereby
> controlling  access to GPUs.

William,

That's why we are planning to support device whitelisting using the
Device Whitelist Controller:

http://www.kernel.org/doc/Documentation/cgroups/devices.txt

Besides GPU devices, there is also the use case for whitelisting
InfiniBand interfaces.


> I'm currently setting up $TMPDIR as a separate tmpfs file system for
> each job and am chgrping that to the job's associated group
> in the process.

In cgroups, there is also the namespace controller for this purpose.


> In general access control for anything  that looks like a file and
> under unix it is alleged "everything is a file".

I can see the flexiblity it brings by combining the extra GID & the
"everything is a file" philosophy.

Our cgroup integration code should be able to add an extra & unique
GID to every job, but IMO by default Grid Engine should not attach
this extra GID as it can affect applications that want to resolve GIDs
- so most likely we will have an option to enable this.

Rayson



>
>
> William

_______________________________________________
users mailing list
[email protected]
https://gridengine.org/mailman/listinfo/users

Reply via email to