Revisiting CPU Performance Scaling for Energy-Efficient Packet Processing Applications

A frequency scaling governor controls the clock speed of CPUs, thereby controlling the performance and power consumption of a CPU.For servers dedicated to packet processing applications, existing scaling governors are sub-optimal in energy efficiency and can benefit from optimization based on application-level metrics such as network loads, enabling more precise control.We propose Puppis, a framework that adapts the CPU core configuration according to network loads based on the system's profiled performance.By selecting the appropriate configuration of CPU cores and CPU frequencies, and steering packets accordingly, Puppis reduces power consumption while satisfying the same throughput requirement and delivering competitive latency performance.We prototype Puppis using the XDP framework.Evaluations with an intrusion prevention system show that power consumption can be reduced by up to 34.1 % compared to existing scaling governors.