Rethinking Software Network Data Planes in the Era of Microservices

Abstract

With the advent of Software Defined Networks (SDN) and Network Functions Virtualization (NFV), software started playing a crucial role in the computer network architectures, with the end-hosts representing natural enforcement points for core network functionalities that go beyond simple switching and routing. Recently, there has been a definite shift in the paradigms used to develop and deploy server applications in favor of microservices, which has also brought a visible change in the type and requirements of network functionalities deployed across the data center. Network applications should be able to continuously adapt to the runtime behavior of cloud-native applications, which might regularly change or be scheduled by an orchestrator, or easily interact with existing “native” applications by leveraging kernel functionalities - all of this without sacrificing performance or flexibility. In this dissertation, we explore the design space of software packet processing applications within the new “cloud-native” era, and we propose a novel paradigm to design, run, and manage software network functions that follow the same approach of micro-services. We present Polycube, a software framework that enables the creation of efficient, modular, and dynamically reconfigurable in-kernel networking components available with vanilla Linux. Polycube exploits the extended Berkeley Packet Filter (eBPF) framework to execute the data plane of those network functions and introduces a set of additional components and common APIs that make it easier to develop and manage those services. We design and evaluate the use of this paradigm through bpf-iptables, a clone of iptables characterized by improved performance and scalability. Then, we explore the possibility of enhancing the capabilities of end-hosts through the use of programmable network interface cards (SmartNICs) to offload partially (of fully) existing packet processing applications, in particular in the domain of DDoS Mitigation. In the last part of the dissertation, we present Kecleon, a compiler framework that can be used to dynamically optimize generic software data planes, taking into account the runtime characteristics and packet processing behavior of the original network function. We believe that the combination of these works can lay the foundation for a new model of packet processing applications that is better suited for modern cloud environments, having the capability to be dynamically re-combined, re-generated, and re-optimized without sacrificing programmability, extensibility and performance.

Related