Nodes with GPU are expensive and it's likely a bunch of pods
using them will be created in a batch. In this case we can
wait a bit for all pods to be created to make more efficient
scale-up decision.
When a machine with GPU becomes ready it can take
up to 15 minutes before it reports that GPU is allocatable.
This can cause Cluster Autoscaler to trigger a second
unnecessary scale up.
The workaround sets allocatable to capacity for GPU so that
a node that waits for GPUs to become ready to use will be
considered as a place where pods requesting GPUs can be
scheduled.
Bartłomiej Wróblewski
Jiaxin Shan
Łukasz Osipiuk
Maciej Pytel
Beata Skiba