Realizing High Performance NFV Service Chains

Network functions (NFs) hold a key role in networks, offering in-network services, such as enhanced performance, policy enforcement, and security. Traditionally, NFs have been implemented in specialized, thus expensive hardware. To lower the costs of deploying NFs, network operators have adopted network functions virtualization (NFV), by migrating NFs from hardware to software running in commodity servers. Several approaches to NFV have shown that commodity network stacks and drivers (e.g., Linux-based) struggle to keep up with increasing hardware speed. Despite this, popular networking services still rely on these commodity components. Moreover, chaining NFs (also known as service chaining) is challenging due to redundancy in the elements of the chain. This licentiate thesis addresses the performance problems of NFV service chains. The first contribution is a framework that (i) profiles NFV service chains to uncover performance degradation reasons and (ii) leverages the profiler’s data to accelerate these chains, by combining multiplexing of system calls with scheduling strategies. These accelerations improve the cache utilization and thereby the end-to-end latency of chained NFs is reduced by a factor of three. Moreover, the same chains experience a multi-fold latency variance reduction; this result improves the quality of highly-interactive services. The second contribution of this thesis substantially revises the way NFV service chains are realized. NFV service chains are synthesized while eliminating redundant input/output and repeated elements, providing consolidated stateful cross layer packet operations across the chain. This software-based synthesis achieves line-rate 40Gbps throughput for stateful and long service chains. This performance is 8.5x higher than the performance achieved by the software-based state of the art FastClick framework. Experiments with three example Internet Service Provider-level service chains show that this synthesis approach operates at 40Gbps, when the classification of these chains is offloaded to an OpenFlow switch.