Design and Evaluation of Cloud Network Optimization Algorithms
The rapid development of software virtualization solutions have lead to huge innovation in Data Center Networks (DCN), with additional capabilities to perform advanced network functions via software elements. Novel protocols designed in the recent years allow for a large path diversity at the edge server level by means of multipath forwarding protocols and virtual bridging functions. Moreover, increasing features are given to DCN elements that become programmable so as to allow, for instance, turning off and on virtual servers for energy consolidation, migrating virtual machines on demand or automatically based on variation of system and network states to improve performance, etc.
In this scope, this dissertation addresses a number of research questions.
The first research question to which we answer is at which extent and in which situations performing multipath forwarding and virtual bridging in DCNs can be beneficial when performing data center optimizations meeting traffic engineering and energy efficiency goals.
We formally formulate the problem of virtual machine placement aware of server and network states, as well as aware of the capability to perform multipath forwarding and virtual bridging. We propose a heuristic approach for its resolution.
We give many insights, showing in particular that virtual bridging brings high performance gains when traffic engineering is the primary goal, and should be deactivated when energy efficiency becomes important. We also determine that multipath forwarding brings relevant gains only when energy efficiency is the primary goal and virtual bridging is not enabled. In a second contribution our focus moves toward the analysis of the relationship between novel flattened and modular DCN architectures and congestion control protocols. In fact, one of the major concerns in congestion control being the fairness in the offered throughput, the impact of the additional path diversity, brought by the novel DCN architectures and protocols, on the throughput of individual endpoints (servers) and aggregation points (edge switches) was an aspect not clearly addressed in the literature.
Our contribution consists of providing a novel comprehensive mathematical programming formulation of the throughput optimization problem based on the proportional fairness principle of the Transport Control Protocol (TCP), and in an extensive series of experiments conducted following the model. We find how much and why the traffic allocation fairness is impacted by the type of DCN architecture employed and by the adopted TCP variant. Finally, in the third contribution we investigate a novel rising virtualization resource orchestration problem in Network Functions Virtualization (NFV) architectures for carrier networks. We define the rising problem of optimally placing Virtual Network Functions and chaining virtualized network functions in the delivery of carrier network services. We propose a mathematical programming formulation taking into consideration both NFV and traffic engineering costs, and describe a math-heuristic approach. We draw quantitatively and qualitatively many insights on the design of NFV infrastructures across carrier network layers (access, aggregation, core) also depending on the type of virtualized network function.
Defence : 04/24/2015 - 11h - Site Jussieu 25-26/105 Jury members : MAIER Guido (Politecnico di Milano, Italy)
NOGUEIRA Michèle (Univ. Federal do Parana, Brazil)
BOUET Mathieu (Thalès, France)
CHEMOUIL Prosper (Orange labs, France)
MEDHI Deep (Univ. of Missouri-Kansas City, USA)
PUJOLLE Guy (Univ. Pierre et Marie Curie, France)
SECCI Stefano (Univ. Pierre et Marie Curie, France)
D. Belabed, S. Secci, G. Pujolle, D. Medhi : “On Traffic Fairness in Data Center Fabrics”, IEEE 3rd International Conference on Cloud Networking (CloudNet 2014), Luxembourg, Luxembourg, pp. 40-45, (IEEE) (2014)