Accelerating Network Functions using Reconfigurable Hardware. Design and Validation of High Throughput and Low Latency Network Functions at the Access Edge

Providing Internet access to billions of people worldwide is one of the main technical challenges in the current decade. The Internet access edge connects each residential and mobile subscriber to this network and ensures a certain Quality of Service (QoS). However, the implementation of access edge functionality challenges Internet service providers: First, a good QoS must be provided to the subscribers, for example, high throughput and low latency. Second, the quick rollout of new technologies and functionality demands flexible configuration and programming possibilities of the network components; for example, the support of novel, use-case-specific network protocols. The functionality scope of an Internet access edge requires the use of programming concepts, such as Network Functions Virtualization (NFV). The drawback of NFV-based network functions is a significantly lowered resource efficiency due to the execution as software, commonly resulting in a lowered QoS compared to rigid hardware solutions. The usage of programmable hardware accelerators, named NFV offloading, helps to improve the QoS and flexibility of network function implementations. In this thesis, we design network functions on programmable hardware to improve the QoS and flexibility. First, we introduce the host bypassing concept for improved integration of hardware accelerators in computer systems, for example, in 5G radio access networks. This novel concept bypasses the system’s main memory and enables direct connectivity between the accelerator and network interface card. Our evaluations show an improved throughput and significantly lowered latency jitter for the presented approach. Second, we analyze different programmable hardware technologies for hardware-accelerated Internet subscriber handling, including three P4-programmable platforms and FPGAs. Our results demonstrate that all approaches have excellent performance and are suitable for Internet access creation. We present a fully-fledged User Plane Function (UPF) designed upon these concepts and test it in an end-to-end 5G standalone network as part of this contribution. Third, we analyze and demonstrate the usability of Active Queue Management (AQM) algorithms on programmable hardware as an expansion to the access edge. We show the feasibility of the CoDel AQM algorithm and discuss the challenges and constraints to be considered when limited hardware is used. The results show significant improvements in the QoS when the AQM algorithm is deployed on hardware. Last, we focus on network function benchmarking, which is crucial for understanding the behavior of implementations and their optimization, e.g., Internet access creation. For this, we introduce the load generation and measurement framework P4STA, benefiting from flexible software-based load generation and hardware-assisted measuring. Utilizing programmable network switches, we achieve a nanosecond time accuracy while generating test loads up to the available Ethernet link speed