This is the accepted version of a paper published in IEEE Communications Surveys and Tutorials. This paper has been peer-reviewed but does not include the final publisher proofcorrections or journal pagination.

Abstract-This paper investigates OpenFlow performance, focusing on how the delay between switch and OpenFlow controller can impact the performance of a network. We use open-source controllers that support network virtualization to evaluate how such delay impacts in ICMP, TCP and UDP traffic. We compare the controllers' flow set-up strategies and we propose several experiments to compare their TCP and UDP performance. In addition, we introduce a new metric to measure UDP packet losses at the beginning of the flow. The results of the measurements indicate that there are large differences in performance between controllers, and that performance depends on switch-controller delay and flow set-up strategy.

Abstract-Network virtualization has been an important research topic for many years but still suffers from the lack of an abstraction level like the one present in virtualization of computing and storage. Our work in progress presented here proposes an architecture for such a network virtualization abstraction. It is deployed as a library, similar to libvirt in computer virtualization, with a unified interface towards the underlying network specific drivers. The architecture will allow management tools to be independent of the underlying technologies. In addition, it will enable programmatic and on-demand creation of virtual networks. A common set of calls is defined to instantiate different virtual networks, using a single node view to provide the user with a suitable abstraction of the network. We describe a prototype of our proposed architecture on top of an OpenFlow-enabled network. We demonstrate its feasibility for creating isolated virtual networks in a programmatic and on demand fashion.

Priority-based scheduling policies are commonly used to guarantee that requests submitted to the different service classes offered by cloud providers achieve the desired Quality of Service (QoS). However, the QoS delivered during resource contention periods may be unfair on certain requests. In particular, lower priority requests may have their resources preempted to accommodate resources associated with higher priority ones, even if the actual QoS delivered to the latter is above the desired level, while the former is underserved. Also, competing requests with the same priority may experience quite different QoS, since some of them may have their resources preempted, while others do not. In this paper we present a new scheduling policy that is driven by the QoS promised to individual requests. Benefits of using the QoS-driven policy are twofold: it maintains the QoS of each request as high as possible, considering their QoS targets and available resources; and it minimizes the variance of the QoS delivered to requests of the same class, promoting fairness. We used simulation experiments fed with traces from a production system to compare the QoS-driven policy with a state-of-the-practice priority-based one. In general, the QoS-driven policy delivers a better service than the priority-based one. Moreover, the equity of the QoS delivered to requests of the same class is much higher when the QoS-driven policy is used, particularly when not all requests get the promised QoS, which is the most important scenario. Finally, based on the current practice of large public cloud providers, our results show that penalties incurred by the priority-based scheduler in the scenarios studied can be, on average, as much as 193% higher than those incurred by the QoS-driven one.

No abstract