DPDK-based systems enable high-performance packet processing, yet their energy behavior remains not fully understood: developers cannot attribute energy consumption to individual dataplane events. Existing tools expose either fine-grained execution traces or coarse-grained power metrics, but fail to correlate them. We present EnergyTracer, a cross-layer observability framework that directly correlates packet-processing events with system-wide energy consumption. It combines low-overhead DPDK fast-path tracing with kernel CPU power-state monitoring and hardware energy measurements, synchronized into a unified timeline, enabling millisecond-scale per-event energy attribution with minimal impact on data plane performance.Evaluated on a 5G gNB workload, EnergyTracer incurs minimal overhead (20 CPU cycles per event) and reveals hidden relationships between bursty traffic, CPU power states, and energy usage, enabling energy-aware optimization of packet-processing systems.

Power efficiency is crucial in softwarized 5G gNBs due to rising energy costs and CO_2 emissions from the networking infrastructure. SmartNICs have the potential to accelerate packet processing tasks. However, their limited resources and functionalities make handling complex tasks like retransmissions challenging. Offloading operations to server-class processors can improve scalability but also increase power consumption, warranting careful management. This study examines the performance and power consumption of splitting the processing of the 5G gNB between SmartNICs and DPDK-enabled servers. We implement an adaptive CPU power conserving strategy in the DPDK data plane that dynamically adjusts CPU power states using P-states and C-states based on traffic load. Our evaluation shows that when using five SmartNICs, we can achieve 95 MPPS, demonstrating that offloading can boost throughput by up to 40% for gNB header processing. Our adaptive power management limits CPU power consumption growth to a maximum of 11%, compared to 52% without such management, balancing performance and power efficiency effectively.

Power efficiency is a growing concern in softwarized 5G gNBs due to increasing energy costs and carbon emissions. While SmartNICs offer low-power acceleration for packet processing, they struggle with complex operations, often requiring offloading to servers, which raises power usage. This study evaluates the performance and power consumption when dividing 5G gNB data plane tasks between SmartNICs and a DPDK-enabled host, comparing flow-based and function-based task allocation methods. We introduce an adaptive CPU power management strategy that adjusts CPU power states based on traffic. Results show that deploying five SmartNICs with function-based partitioning, which retains packet buffering on the host and offloads header decapsulation, insertion, and lookup-table operations to the SmartNICs, delivers a throughput of 109 MPPS, which is 173% more than a SmartNIC-only setup and reduces latency by 70% compared with a host-only setup. Adaptive power management lowers total power consumption by 12.5% in the optimal partitioning while preserving high throughput and low latency.

The traditional method of data plane programming involves deploying a single P4 program to a single target. However, different targets have varying capabilities, functionalities, and support for various programming languages beyond P4. Therefore, disaggregating a single data plane program into multiple subprograms that run on different targets can allow us to leverage the strengths of each target, which becomes particularly important in the context of 5G, where some data plane processing functions, such as buffering and retransmission for RLC processing, cannot be effectively expressed in P4. This paper delves into the decomposition of a 5G gNB across a P4-programmable SmartNIC and an x86 server using DPDK-based processing, thus harnessing the strengths of each target. Our evaluation revealed that offloading certain processing to an x86 server can improve throughput by up to 50%, thanks to the scalability of DPDK applications' performance with the number of CPU cores. However, offloading does introduce a slight increase in latency, so the approach should be adjusted based on the specific needs and available resources.

The typical approach to data plane programming involves deploying a single P4 program to a single target. However, different targets have different capabilities, functionalities, and support for various programming languages apart from P4. Consequently, disaggregating a single data plane program into multiple subprograms that run on different targets can take advantage of the strengths of each target, which is particularly important in the context of 5G, as certain data plane processing functions, like buffering and retransmission for RLC processing, cannot effectively be expressed in P4. This paper explores the disaggregation of a 5G gNB across a P4-programmable SmartNIC and an x86 server using DPDK-based processing, leveraging the strengths of each target. We assess the performance of our hybrid approach by varying which parts of the pipeline run on the SmartNIC and the x86, as well as the number of cores allocated on the host for the non-P4 part of the pipeline.