Thanks for your time.
On 2021/9/10 5:23 PM, Michal Koutný wrote:
On Fri, Sep 10, 2021 at 01:30:46PM +0800, brookxu <brookxu.cn@xxxxxxxxx> wrote:
I am a bit confused here. For misc_cgroup, we can only be rejected when the count
touch Limit, but there may be other more reasons for other subsystems.
Sorry, I wasn't clear about that -- the failures I meant to be counted
here were only the ones caused by (an ancestor) limit. Maybe there's a
better naem for that.
Therefore, when we are rejected, does it mean that we have touch
Limit? If so, do we still need to distinguish between max and fail?
(for misc_cgroup)
r
`- c1
`- c2.max
`- c3
`- c4.max
`- task t
`- c5
Assuming c2.max < c4.max, when a task t calls try_charge and it fails
because of c2.max, then the 'max' event is counted to c2 (telling that
the limit is perhaps low) and the 'fail' event is counted to c4 (telling
you where the troubles originated). That is my idea. Although in the
case of short-lived cgroups, you'd likely only get the hierarchically
aggregated 'fail' events from c3 or higher with lower (spatial)
precision.
What would be the type of information useful for your troubleshooting?
Through events and events.local, we can determine which node has
insufficient resources. For example, when the ‘events’ is large, we
traverse down and use events.local to determine which node has
insufficient resources. 'fail' counter does not seem to provide more
effective information in this regard. When 'fail' is big, it seems that
we still need to use events and events.local to determine the node of
insufficient resources. I am not very sure what details can we learn
through 'fail' counter.
Cheers,
Michal
