下世代光數據中心網路之關鍵核心技術-總計畫及子計畫一：下世代光數據中心網路之架構設計及模型化

Abstract

數據中心網路(DCN)被視為支援未來雲端運算及興起中之分散式計算的最重要技術之一。下世代數據中心網路最根本的挑戰是如何將指數成長的巨量(如：數十萬部)伺服器以有效率且低成本的方式連接在一起，同時滿足數項重要的條件：高可擴展性、高吞吐量、高容錯力、低延遲、無封包遺失及低功率消耗。由於有限的吞吐量及高功率消耗，目前使用電訊號的數據中心網路已經成為數據中心其持續成長的運算能力之主要技術瓶頸。同時，過去二十年來光網路技術在長途、都會及區域/存取網路領域中已成為連結通訊節點的卓越技術。眾多研究人員投注大量心力於研究如何使用光通訊及光交換機技術實現數據中心網路，也就是所謂的光數據中心網路。所幸近來光元件(如：光/電/光轉換器、被動陣列波導光柵交換器)的進步，光數據中心網路之低封包遺失、低功率耗散、低互相干擾及極高的吞吐量，都確保其能長遠滿足上述數項數據中心網路的需求。 對光數據中心網路而言，為滿足數據中心網路的數項條件，必須先發展數項關鍵核心技術，其中包含：(1)數據中心及光交換機之架構與模型建構、(2)平行化排程及其最佳化、(3)低功耗波長轉換技術、(4)高容量點對點光通訊技術，以上四項研究被統整於該整合型計畫之中。在此(技術1)之三年計畫中，我們共有兩個研究主軸。在第一個研究主軸，我們關注於數據中心網路架構的設計及模型建構；傳統的Fat-Tree數據中心網路架構為了克服資料傳輸的瓶頸，必須在樹的根部使用昂貴的超高容量交換器，造成在支援巨量伺服器時的擴展性問題；此外，雖然Fat-Tree在理論上能夠於任意兩個伺服器間都提供高頻寬，但於此種多分支且流量難以預測的網路中提供負載平衡卻是極大的挑戰。我們的計畫目標是設計並建構一全新且可行的數據中心網路架構，以提供高吞吐量(低延遲)同時能使負載平衡、高可擴展性但無效能瓶頸、並且在低流量時動態調整網路拓墣以達到最低功耗。第二個研究主軸則是設計、分析一個在數據中心網路中使用的高效能光交換機系統並製作其原型。在此之前，我們曾設計過一個10Gb/s/wavelength的分波長多工光封包交換機系統，雖然此二系統之目的都是提供高吞吐量，但在該計畫中的數據中心光交換機系統與之前的分波長多工光封包交換機系統有兩個方面的不同：一是由於光/電/光轉換器逐漸成熟，此數據中心光交換機系統將會採用電緩衝器(可能額外加上數個光纖延遲線光緩衝器)；其二是此光交換機系統必須滿足數據中心網路的低功耗要求。主要的設計策略將包含功率控制/管理及使用共用波長轉換器。最後，此三年內要完成的各項工作總結於以下roadmap中。

Data center networking (DCN) has been envisioned as one of the most prominent technologies for supporting future cloud computing and ever-increasing distributed computing applications. A fundamental challenge in next-generation DCN is how to efficiently and cost-effectively interconnect an exponentially increasing number (e.g., hundreds of thousands) of servers while simultaneously meeting a multitude of requirements. The crucial requirements include high scalability, throughput, fault tolerance, low latency, no loss, and low power consumption. Due to limited throughput and large power consumption, electrical-based data center networking has become a technological bottleneck and a dominant factor limiting the continued scaling of processor performance. Meanwhile, optical networking has been a preeminent technology for connecting telecommunicating nodes in the long haul, metro, and local/access geographic areas over two decades. Together, researchers have recently drawn tremendous attention and interests to the realization of DCN by means of optical transmission and switching technologies, so called optical-based DCN. Thanks to recent advances in photonic devices (e.g., E/O/E devices, and passive Array Waveguide Grating (AWG) switches), such optical-based DCN has been shown capable of providing long-term solutions to the above limitations with low loss, low power dissipation, low interference, and exceedingly high throughput. For optical-based DCN, to satisfy the DCN requirements, several key enabling technologies must first be developed which involves investigating the following four research tasks (undertaken by this integrated project): (1) data-center and optical-switch architectures and modeling; (2) parallel scheduling and optimization; (3) low-power-consumption wavelength converter techniques; and (4) high-capacity point-to-point optical interconnect. In this three-year project (task one), we focus on two lines of research work. In the first line of work, we focus on the design and modeling of data-center architectures. The traditional fat-tree DCN architecture, to overcome bottlenecks, requires the use of expensive exceedingly high-capacity switches toward the root of the tree, resulting in a severe scalability problem when supporting an immense number of servers. Besides, although the fat-tree network can theoretically support high bandwidth between any pair of servers, unfortunately, balancing network traffic in such a network with highly divergent and unpredictable traffic matrices poses significant challenges. Our aim of the project is to design/model viable and novel architectures achieving high throughput (and low latency) with load balancing, high scalability with no performance bottleneck, and dynamic adjustment of network topology subject to minimize power consumption under low traffic demands. In the second line of work, we aim at the design, analysis and prototype of a high-performance data-center optical switching system (DOSS). We have earlier prototyped a 10-Gb/s/wavelength WDM optical packet switching system. While both systems are aimed to achieve high throughput, the data-center optical switching system in this project differs from the 10-Gb/s system on two perspectives. First, due to the maturity of O/E/O devices, electrical-buffering will be adopted in this optical switching system for DCN (possibly in addition to a handful of FDL-based optical buffer). Second, the optical switching system must be designed to meet the low power consumption requirement for DCN. Design strategies include power control/management schemes and sharing of wavelength converters. Finally, the tasks to be accomplished in these three years are summarized in the following roadmap.