As datacenter networks evolve toward ultra-high-speed links, the energy footprint of host-side packet processing grows increasingly significant. While prior works have scrutinized the performance overhead of networking stacks, the interaction between energy efficiency mechanisms and networking remains underexplored. This paper presents a systematic study of the impact of CPU frequency scaling - both core and uncore - on host networking performance and power consumption. We analyze two dominant communication primitives in modern datacenters: the Linux TCP/IP stack and RoCEv2 RDMA. Our findings reveal that uncore components operate at excessively high frequencies, contributing to overall system power consumption substantially, and that recent Linux's core scaling governors make suboptimal decisions regarding target frequency. By bounding uncore frequency, we achieve up to 46.1% power savings with lower degree of throughput degradation (<11.6%) for TCP-based key-value store workloads. The significance of uncore also holds for RDMA: its offloaded design minimizes responder CPU activity, yet high bandwidth workloads still trigger uncore power draw.
Ok et al. (Wed,) studied this question.