Horizontal and vertical scaling frameworks for network function virtualization

Abstract

NFV (Network function virtualizations) promises the benefit of the virtual environment -- elastic scaling, fault tolerance, rapid deployment of new services -- by moving network functions from fixed-function middlebox infrastructures to software on virtual machines run on commodity servers. One of the challenges arise from the NFV (Network Function Virtualization) is that the gap between the promise of network function virtualization (NFV) and the practice as is. My dissertation focuses on the scalability, one of such gap

processing throughput should increase linearly with added resources (CPU cores, VMs, processes) without imposing much overhead. However, building scalable network functions is difficult since they are stateful

Some state needs to be migrated to other instance (VM, core, process) to preserve locality as network traffic is re-balanced (e.g., per-flow structures), while others need to be shared across instances (e.g., aggregated counters, a list of detected hosts). The statefulness is a source of performance overhead such as core/machine-crossing overhead, locking for shared structures as well as making the programming task complex. In this dissertation, we identify the challenges for achieving a scalable performance of network functions and propose three systems for them. Firstly, based on the observation of the state access patterns of network functions, we build a network function building programming model and framework to support elastic horizontal scalability (S6). It shows linear throughput scaling with 2-3 orders of magnitude lower per-packet added latency than existing state-of-the-art systems. Secondly, we present a hardware assisted flow load-balancing mechanism (Symmetric RSS) to avoid core-crossing overhead and implement a high-performance and multi-core scalable flow monitoring system (MonBot) with the mechanism. Using the system, we conduct large-scale real-time measurement study on the live 3G commercial traffic. Lastly, we design and implement a multi-core scalable network stack (mTCP) to overcome the current kernel's inefficiency. It shows 25 times better connection set-up performance than Linux kernel.