kube-proxy cannot set the conntrack hashsize inside a container

[twelho]

When running make integration-tests-container, kube-proxy in node0 tries to update /sys/module/nf_conntrack/parameters/hashsize in case it is not large enough, but this fails since it can only be updated from the root netns (on host). This causes the integration test to fail, the first node never completes initialization.

This issue seems to only show up with high nproc counts, tested on a system with 24 hardware threads. Updating the hashsize value requires root, so maybe instead of making wksctl directly write it, we could have a preflight check?

• Same issue in kube-spawn: kube-proxy stuck in CrashLoopBackOff kinvolk/kube-spawn#14
• More info (e.g. how the value is calculated): http://blog.michali.net/category/kubernetes/page/2/ (search for "hashsize")