Orchestration Efficiency

The Orchestration Efficiency (OrchestrationE) is defined in the context of a multi-GPU system as the ratio between the non-idle time, which includes useful time and time pertaining to data transfers, for the GPU where the sum of both is the highest, and the total execution time.

\[OrchestrationE = \frac{Max(DeviceUseful + DeviceWaitForData)}{Runtime}\]

This metric represents the portion of the runtime the GPUs are being actively used. A low value can be attributed to waiting for work from the host or dependencies between kernels.

In order to fully understand the formulas, you may also visit the glossary of the metrics terms.