They shouldn't be coordinating often unless the benchmark is poorly suited.

Only for supercomputer tasks like weather forecasting. Not for near-embarrassingly parallel jobs like high traffic web servers and render farms.