Optical Interconnects in Data Center Networks

Abstract

Optical Interconnects in Data Center Networks Emerging cloud computing and web-based applications have created a need for powerful data centers. Optical interconnects are replacing the electrical counterparts due to their benefits. Electrical interconnects have various problems like energy consumption, fixed throughput, and limited reconfigurability. The thesis evaluates various optical interconnects in literature and analyzes various aspects of optical interconnects in data center networks. The conclusions of various analyses are used to propose architectures which give several advantages over existing architectures. Main contributions of the thesis are twofold, one: analysis of existing architectures is carried out for the first time in terms of blocking probability, signal degradation, port count, computational complexity of routing algorithm and input and output queueing analysis and two: four new architectures are presented that give benefits of very low packet loss, large port count, reduce computational complexity regardless of port count and contention less behavior under various traffic patterns. The blocking probability analysis involves the analysis of the effect of multiplexing techniques and optical components. The signal degradation analysis involves the analysis of various optical components which affect the signal degradation and ways to overcome signal degradation have been identified. The port count analysis involves the analysis of the limitations of port count of different architectures based on optical components and multiplexing techniques. The computational complexity analysis involves the analysis of the complexity of routing algorithms of the architectures. Various architectures are proposed based on the analysis of existing architectures. The input and output queueing analysis of the existing architectures and proposed architectures is performed by changing the size of the output queue buffer. The affect of the output queue buffer size is analyzed on delay and throughput by changing the load for the architectures. Maximum throughput is also analyzed by changing the buffer size of output queue. The input and output queueing models of architectures are also analyzed by simulation. A contention less architecture is proposed which can reduce contention under various traffic patterns. It exploits optical wavelength, code and space multiplexing. It eliminates the need for electronic buffering, which reduces the latency. It is robust to all-to-one and many-to-many communication traffic patterns. A very low packet loss architecture is proposed. This architecture can show a certain guaranteed data rate between communicating servers. This architecture is flexible and can be made suitable for a broad range of data center applications, by configuring it. An architecture with very large port count is presented. This architecture has large port count compared to other architectures. This architecture also has low energy consumption due to use of passive optical components. It is suitable for large scale data center networks. It is also discussed how various optical components and multiplexing techniques affect the overall port count of optical interconnects. An architecture with low computational complexity of routing algorithm is proposed. If a control algorithm is more complex it requires more hardware for implementation which causes increase in latency, increase in cost and increased difficulty in implementation. The computational complexity of proposed architecture is not affected by the number of nodes. This architecture is suitable for low latency and real time applications. This architecture exploits three dimensions of multiplexing namely; the optical wavelength division multiplexing, optical code division multiplexing, and optical space division multiplexing.