Virtualization overhead
A virtualization overhead model determines how much performance a GPU loses when it is shared by virtual machines. The
overhead lowers the capacity the GPU can supply. It is set with the virtualizationOverhead field of a
GPU. The default is none.
The type selects the model:
| Type | Fields | Behaviour |
|---|---|---|
none | No overhead, as with GPU pass-through. This is the default. | |
constant | percentageOverhead | A fixed overhead, as a fraction of the capacity. When percentageOverhead is omitted, 0.05 (5%) is used. |
shareBased | An overhead that depends on the number of virtual machines sharing the GPU, based on Garg et al., Empirical Analysis of Hardware-Assisted GPU Virtualization. |
Example
"virtualizationOverhead": {
"type": "constant",
"percentageOverhead": 0.25
}