Performance, Isolation and Service Guarantees in Virtualized Network Functions
Abstract: A network is generally a collection of different hardware-based network devices carrying out various network functions, (NF). These NF implementations are special purpose and expensive. Network function virtualization (NFV) is an alternative which uses software-based implementation of NFs in inexpensive commodity servers. However, it is challenging to achieve high networking performance due to bottlenecks in software, particularly in a virtualized environment where NFs are implemented inside the virtual machines (VM). The performance isolation is yet another challenge, which means that the load on one VM should not affect the performance of other VMs. However, it is difficult to provide performance isolation due to resource contention in a commodity server. Furthermore, different NFs may require different service guarantees which are difficult to ensure due to the non-deterministic performance behavior of a commodity server.In this thesis we investigate how the challenges of performance, isolation and service guarantees can be addressed for virtual routers (VR), as an example of a virtualized NF. It is argued that the forwarding path of a VR can be modified in an efficient manner in order to improve the forwarding performance. When it comes to performance isolation, poor isolation is observed due to shared network queues and CPU sharing among VRs. We propose a design with SR-IOV, which allows reserving a network queue and CPU core for each VR. As a result, the resource contention is reduced and strong performance isolation is achieved. Finally, it is investigated how average throughput and bounded packet delay can be guaranteed to VRs. We argue that a classic rate-controlled service discipline can be adapted in a virtual environment to achieve service guarantees. We demonstrate that firm service guarantees can be achieved with little overhead of adding token bucket regulator in the forwarding path of a VR.
CLICK HERE TO DOWNLOAD THE WHOLE DISSERTATION. (in PDF format)