Enhancing programmability for adaptive resource management in next generation data centre networks

QA75 Electronic computers. Computer science

Jouet, Simon ; Pezaros, Dimitrios

University:University of Glasgow

Department:School of Computing Science

关键词: SDN, data centre, data plane programmability, next generation networks, BPFabric, OTCP.;   

Others  :  http://theses.gla.ac.uk/8535/1/2017JouetPhD.pdf

来源: University of Glasgow

【 摘 要 】

Recently, Data Centre (DC) infrastructures have been growing rapidly to support a wide range of emerging services, and provide the underlying connectivity and compute resources that facilitate the "*-as-a-Service" model. This has led to the deployment of a multitude of services multiplexed over few, very large-scale centralised infrastructures. In order to cope with the ebb and flow of users, services and traffic, infrastructures have been provisioned for peak-demand resulting in the average utilisation of resources to be low. This overprovisionning has been further motivated by the complexity in predicting traffic demands over diverse timescales and the stringent economic impact of outages. At the same time, the emergence of Software Defined Networking (SDN), is offering new means to monitor and manage the network infrastructure to address this underutilisation.This dissertation aims to show how measurement-based resource management can improve performance and resource utilisation by adaptively tuning the infrastructure to the changing operating conditions. To achieve this dynamicity, the infrastructure must be able to centrally monitor, notify and react based on the current operating state, from per-packet dynamics to longstanding traffic trends and topological changes. However, the management and orchestration abilities of current SDN realisations is too limiting and must evolve for next generation networks. The current focus has been on logically centralising the routing and forwarding decisions. However, in order to achieve the necessaryfine-grained insight, the data plane of the individual device must be programmable to collect and disseminate the metrics of interest.The results of this work demonstrates that a logically centralised controller can dynamically collect and measure network operating metrics to subsequently compute and disseminate fine-tuned environment-specific settings. They show how this approach can prevent TCP throughput incast collapse and improve TCP performance by an order of magnitude for partition-aggregate traffic patterns. Futhermore, the paradigm is generalised to show the benefits for other services widely used in DCs such as, e.g, routing, telemetry, and security.