https://hgpu.org/?p=17944
Improving Communication Performance in GPU-Accelerated HPC Clusters