云计算数据中心高扩展全光互连架构设计

Due to the attractive strengths such as high bandwidth, low power consumption, transparent transmission, optical interconnection technology is a promising way to solve the bottleneck of the bandwidth and power consumption of the conventional electrical interconnects in data center. However, prior optical data center networks suffer from some the following issues, including coarse switching granularity, high control overhead, low scalability and poor throughput. To solve above problems, this project sets goals of building a new optical data center network with low blocking ratio, high scalability, large capacity, and low power consumption. By Leveraging the theory of composite graph and constraints of optical propagation properties, a new network topology is designed to optically connect tens of thousands servers. Then by exploiting the bi-directional transparency of optical devices, further combing with the specific connection rules, a distributed control protocols, including switching strategy with all-optical feedback operation, forwarding-port based source routing algorithm, and collision domain aware wavelength allocation mechanism, are designed to ensure the low interaction overhead, lossless packet forwarding, and multiple granularity of message operation. Based on the latest development of silicon photonics technology and the method of designing non-blocking switching structure, an optical switch, enabling the fast configuration, multi-domain switching, and.transparent transmission is proposed to well support the communication protocols. The research achievements of this project will provide a typical application scenario, which takes benefit from the latest research progresses of silicon photonics. It also solves the contradictory demands for the bandwidth and power consumption. Overall, the new design will provides the low-energy and high-performance optical interconnects for large-scale data center.

由于具备高带宽、低能耗、透明传输等特性，光互连被认为是解决云计算数据中心电互连网络能耗和带宽瓶颈的有效技术。针对现有光互连方案在交换粒度、控制开销、互连规模、有效吞吐等方面的不足，本课题以设计低阻塞、易扩展、低能耗、高容量光互连网络为目标，通过引入复合图理论和基于光传输特性的约束方法设计拓扑结构，实现数万节点的低能耗全光互连；利用光器件双向传输特性，结合特定拓扑的连接规则设计分布式控制协议，包括基于全光反馈的光交换策略、基于转发端口的源路由算法、基于冲突域抑制的波长分配方案等，实现低开销、无丢包、多粒度的信息传输控制；基于硅光技术的最新进展和无阻塞交叉开关的构建方法，设计快速配置、多域（时域/空域/波长域）操作、信道透明的光交换节点，实现对通信策略的有效支持。课题的研究将为硅光技术最新成果提供典型应用场景，解决带宽增长与能耗约束的矛盾，为大规模数据中心提供一种低能耗、高性能的光互连方案。

由于具备高带宽、低能耗、透明传输等特性，光互连被认为是解决云计算数据中心电互连网络能耗和带宽瓶颈的有效技术。针对现有光互连方案在交换粒度、控制开销、互连规模、有效吞吐等方面的不足，本课题以设计低阻塞、易扩展、低能耗、高容量光互连网络为目标，在网络拓扑设计、通信策略设计及交换节点设计等方面展开研究。.依托项目，设计了可重构光电混合互连网络架构。该网络通过可扩展的拓扑结构解决传统光互连网络无法大规模扩展的问题，通过高容量光交换节点实现网络层级的压缩，从而降低网络通信的平均时延。进一步，根据分布式计算业务的流量特点设计光链路调度算法，实现应用通信需求的快速发现以及通信逻辑拓扑与物理网络的高效适配，同时有效降低调度算法的复杂度和决策时延。为解决连通性问题，课题组提出了中继路由的策略，有效提升了网络的连通性及光电交换机的协同能力。仿真及实验表明，可重构光电混合互连网络能够在节约设备数量及线缆开销的情况下，实现对业务的高效承载。在部分流量模式下，该网络能够达到理想无阻塞网络的性能。.针对现有数据中心光交换网络控制平面存在流量收集开销大、拓扑配置算法复杂、配置周期长等问题。对此，课题组开发了光交换网络快速控制平面，在流量收集、拓扑计算、光交换机配置、光电协同等环节提出了优化设计。通过实物测试，光交换平面的控制时延约46ms，相比经典光网络控制系统，该控制系统的控制时延降低了3倍以上。.上述研究有效验证了光互连技术在计算网络的潜在价值，对我国数据中心网络、高性能计算网络和分布式机器学习网络的技术革新及技术领先具有积极推动作用。