Energy-efficient Computing on Distributed GPUs using Dynamic Parallelism and GPU-controlled Communication