contrib: add example for enabling per-container RDT monitoring #228
Conversation
marquiz
force-pushed
the
devel/rdt-monitoring-kustomize
branch
from
a074068 to
7864431
Compare
| delete() { | ||
| # MON group is reaped as part of the CLOS (by the runtime) if it was under | ||
| # a dedicated CLOS created for this container | ||
| if [ "$closid" != "$id" ]; then |
There was a problem hiding this comment.
$closid ? Shouldn't this be [ "$clos" != "$id" ] ?
There was a problem hiding this comment.
YES, well-spotted, thanks. Fixed
Signed-off-by: Markus Lehtonen <markus.lehtonen@intel.com>
marquiz
force-pushed
the
devel/rdt-monitoring-kustomize
branch
from
7864431 to
91fbf06
Compare
This is using the hook-injector as a dependency, see the kustomization.yaml ...
resources:
- ../../hook-injector/unstableIow, this takes the hook-injector sample deployment as a base and does customization on top of that. |
|
Hi team! Got a reference to this PR from @kad . This is a nice, simple approach! 👍 I'm working on an NRI-based container monitor that handles pre-existing containers: unvariance/collector#252 One of the problems is how to reliably pull all tasks of a cgroup into the resctrl group given tasks are live (they can fork as we're adding tasks, creating coverage gaps). If you have any feedback please share. |
@yonch I think there is unfortunately no easy reliable, yet unintrusive way of doing that. A potentially intrusive way is cgroup freezing. Freeze the cgroup of the container (or the whole pod), wait for it to get frozen, assign all the tasks to the resctrl group (if this works while the cgroup/task is frozen, haven't tried it), then unfreeze the cgroup. |
Yes, unfortunately I cannot think of any other race-free approach than the freezer which obviously is a BIG hammer. If you ask me the kernel should provide a simple way to do the migration but 🤷♂️ |
4 participants
No description provided.