OptoCloud: Ultra-fast optically interconnected heterogeneous Data Centers

大多数人类活动，包括交通、互联网、银行、公共卫生和娱乐，都依赖于数据中心。预计云流量将呈指数级增长，占全球流量的95%。2015年，全球数据中心的总功耗高于英国的全国功耗，预计到2030年将增加15倍。目前，所有的数据中心网络都是基于分层电子分组交换网络形成的;然而，它们无法跟上需求，从而在数据增长与摩尔定律之间产生了越来越大的差距。因此，虽然计算节点的功耗（以每秒触发器数为单位）在过去18年中增加了65倍，但节点通信带宽仅增加了4.8倍，而每次触发器通信的字节数减少了8倍。这就创建了一个计算到通信墙，最大限度地减少了数据移动，并限制应用程序在本地运行。此外，这些系统还遭受非常高的中值时延O（100微秒）（100微秒的量级）和99.9百分位尾时延O（100毫秒），从而损害系统和应用性能。OptoCloud奖学金旨在设计和构建节能、经济、可扩展、单跳和纳秒速度的光电路交换网络。这将互连由服务器、CPU、加速器、神经形态处理器、内存元素、存储组成的异构系统，以支持不同部件（机架、行尾）和数据中心大小（中小型~10- 100，000到~ 1，000，000服务器群）。至关重要的是，该网络旨在提供零数据丢失，而没有网络内a）缓冲，B）主动交换和路由，以及c）网络报头寻址和处理，以最小化复杂性，并消耗非常低的功率。此外，系统还将以同步方式固有地支持1对1、1对N、N对N和N对1连接，而不需要用于多播/广播的数据复制，这在当前是不可能的。这是支持各种工作负载的关键，例如存储缓存，大规模数据库查找，训练分布式深度神经网络，使用allreduce，broadcast and reduce，gather and scatter，all-to-all等通信原语的并行计算。为了实现这些目标，OptoCloud将探索亚纳秒光交换、接近无接收器的低功耗收发器和纳秒调度的基本挑战，这些挑战能够在每10 - 100纳秒内重新配置电路并塑造IT和网络拓扑。它的目标是在以下方面提供数量级的改进：a）交换，B）调度和网络拓扑重新配置，c）功耗，d）中延迟和尾延迟，最后e）零数据丢失的吞吐量。PI将与PDRA，博士生，工业合作伙伴合作（微软、Finisar、Xilinx、住友电气），以及大学（哥伦比亚和雅典国立技术大学）并形成一个独特的计算和光网络生态系统，以在方法上回答基本问题，同时反映所有必要的要求提出的概念，并使用工业驱动的用例场景严格评估开发的技术。

The majority of human activities, including transport, Internet, banking, public health and entertainment, depend on Data Centers. Cloud traffic is forecasted to grow exponentially and account for 95% of global traffic. In 2015, the total power consumption of data centers worldwide was higher than the national power consumption of the UK and is predicted to increase up to 15-times by 2030. Currently, all data center networks are formed based on hierarchical electronic packet switched networks; however, they can't keep up with demand creating a ever increasing gap between data growth and Moore's Law. So, while compute node power, measured in flop/s, has increased by 65 times in the last 18 years, the node communication bandwidth has only increased by 4.8 times and the bytes communicated per flop have decreased 8 times. This creates a computation to communication wall, minimizing data movement and constraining applications to operate locally. In addition, these systems also suffer from very high median latencies, O(100microseconds) (order of 100microseconds), and 99.9-percentile tail latencies, O(100ms), to the detriment of the system and application performance.The OptoCloud fellowship aims to design and build an energy efficient, cost effective, scalable, single hop, and nanosecond speed optical circuit switched network. This will interconnect heterogeneous systems made of servers, CPUs, accelerators, neuromorphic processors, memory elements, storage to support different parts (rack, end-of-row) and sizes of data centers (small-medium size ~10-100,000 to ~1,000,000 server farm). Crucially, the network aims to offer zero data loss, without in-network a) buffering, b) active switching and routing, and c) network header addressing and processing to minimize complexity, and to consume very low power. Furthermore, the system also will inherently support 1-to-1, 1-to-N, N-to-N and N-to-1 connectivity in a synchronous manner without the need for data replication for multi/broad -casting, currently not possible. This is key to support diverse workloads such as storage caching, large-scale database lookups, training distributed deep neural networks, parallel computing that use communication primitives such as allreduce, broadcast and reduce, gather and scatter, all-to-all among others. To achieve these, OptoCloud will explore the fundamental challenges of sub-nanosecond optical switching, near receiver-less low-power transceivers and nanosecond scheduling able to reconfigure circuits and shape IT and network topologies every 10s-100s of nanoseconds. It aims to offer orders of magnitude improvement in a) switching, b) scheduling and network topology re-configuration, c) power consumption, d) medium and tail latency and finally e) throughput with zero data loss.The PI will work with the PDRAs, PhD students, industrial partners (Microsoft, Finisar, Xilinx, Sumitomo Electric), as well as universities (Columbia and National Technical University of Athens) and form a unique compute and optical network ecosystem to methodologically answer fundamental questions while reflecting all necessary requirements on the proposed concepts, and rigorously evaluating developed technologies using industrial driven use case scenarios.