problem of resource contention

[SchoIsles]

Hi，
while assign more than three pods to one kubelet node at one moment，caused vf resource contention，report error like “failed to move vf 16 to netns: no such device”
It looks like two pods allocFreeVF got same vf，the second pod try to use device of the vf but already linked into the first pod's ns.
does your production envirment scheduler doesn't consider this question?
and if I need to perfect the plugin by add some lock, or my use method is wrong?

[SchoIsles]

add ex-lock in getVFDeviceName and unlock in setupVF resolved

[rkamudhan]

We didn't facing such issue. Can you explain your deployment model ?

[SchoIsles]

@rkamudhan
kubernetes 1.7.3, create an deployment with nginx image，2 kubelet node(256G memory, 40c)

scala replicas from 2 to 8, every node assign 3 pods, successful created 2, but the 3rd pod's network container failed because of "failed to move vf 16 to netns: no such device", change into the directory named virtfn16/net, there is blank. After I remove the failed container, kubelet recreate it successful

after add the lock, test replicas 0 to 60, all succeed

[hustcat]

I haven't test this, but I think this maybe happen when creating multiple pods at the same time. Thanks @okletswin