Write documentation according to the design document
For gpu_scheduling_behavior; reference Preferential scheduling page.
New page. Should include:
- The settings for gpu_scheduling_behavior.
- Edge case behaviors:
- Non-GPU apps will be scheduled to run on GPU nodes if all of the GPUs are used
- Apps/pods with persistent volumes will continue to launch on the node for which the reservation was made originally; if you change the GPU scheduling behavior from unrestricted to restricted after a non-GPU resident task is scheduled on a GPU resource node, then the resident task will continue to launch on the GPU node.
- A matrix of the Mesos scarcity configuration (see this pull request, there are multiple Mesos configuration parameters), Marathon opt-in feature flag, and Marathon GPU scheduling behavior, to spell out whether or not an app will be launched.
- At this time, we will not document GPU_SCHEDULING_BEHAVIOR=unrestricted label