Powerful GPUs or Fast Interconnects: Analyzing Relational Workloads