On Designing Convertible Data Center Network Architectures
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Most data centers deploy fixed network topologies. This brings difficulties to traffic optimization and network management, because bandwidth locked up in fixed links is not adjustable to traffic needs, and changes of network equipments require cumbersome rewiring of existing links. Moreover, each network topology has unique properties, so it is infeasible to use a one-size-fit-all structure to satisfy the heterogeneous and ever-changing service requirements in data centers. We believe the solution is to build convertible data center network architectures, which can dynamically change the network topology through cable rewiring fully automated by software. We leverage low-cost small port-count converter switches to enable topology change and introduce three convertible data center architectures to experiment with the idea.
OmniSwitch is a production-ready modular container that serves as a universal building block for constructing data centers of various scales. It interleaves converter switches with Ethernet switches to provide local topology optimization and large-scale connectivity. Flat-tree improves transmission performance by dynamically changing topological clustering characteristics of the network. It enables conversion between Clos and approximate random graph networks to provide a suitable network topology for each traffic workload. ShareBackup improves reliability with the concept of "shareable backup". It allows the network to share a small pool of backup switches that can be brought online instantaneously to recover from failures. These works demonstrate the powerful idea of convertible networks, which has the potential to improve a wide range of network performance characteristics, e.g. traffic optimization, load balancing, failure recovery, network expansion, power saving etc.
Description
Advisor
Degree
Type
Keywords
Citation
Xia, Yiting. "On Designing Convertible Data Center Network Architectures." (2018) Diss., Rice University. https://hdl.handle.net/1911/105632.