Distribution policy

A distribution policy determines how the capacity of a CPU or GPU is shared between the tasks on a host. Every host has one for its CPU (cpuDistributionPolicy) and one for its GPU (gpuDistributionPolicy). The default is MAX_MIN_FAIRNESS.

The type selects the policy:

Type	Fields	Behaviour
MAX_MIN_FAIRNESS		Satisfies the smallest demands first, so that no task is starved. This is the default.
BEST_EFFORT	updateIntervalLength (milliseconds, default 1000)	Round-robin over the tasks, rotating every updateIntervalLength, while using as much of the capacity as possible. Inspired by NVIDIA's vGPU scheduler.
EQUAL_SHARE		Gives every task the same share of the capacity, regardless of its demand. Inspired by NVIDIA's vGPU scheduler.
FIRST_FIT		Gives each task its full demand, in order, until the capacity is used up.
FIXED_SHARE	shareRatio (default 1.0)	Gives each active task a fixed fraction of the capacity. Unused shares are not given to other tasks. Inspired by NVIDIA's vGPU scheduler.

Example

"gpuDistributionPolicy": {
    "type": "BEST_EFFORT",
    "updateIntervalLength": 60000
}