The Design Of Novel Data Center Networks And Routing Algorithms

Data centers networks(DCN)are the dominant infrastructure of cloud computing and network applications.They are not only the bridge between large-scale of servers,they are also the basic platform of networking storage and networking computing.Nowadays,more and more severs in DCNs participate in the routing and forward filed,while,more and more switches are used as a computing or storage unit.As a result,computing,storage and the devices are closely related.However,cloud computing still suffers from the bottleneck that comes from the underlying infrastructures.Thus,cloud computing brings DCN fatal chanllenges,among which,the incremental expansion is the most important one.Based on the basic requirement of DCNs,the main work of the paper includes:(1)Surveying the existing DCN topologies systematically.The topological properties of data centers are critical factors that dominate the performance of the whole systems.A number of novel DCN network structures are proposed to improve the network performance.We further propose a taxonomy system to classify such network structures into five categories,i.e.switch-centric data centers,server-centric data centers,modular data centers,random data centers and wireless data centers.(2)Proposing a novel methodology for wired DCNs.We notice that the existing wired DCNs are either structured or random.Both of the two categories own their superiorities,but also suffer from their drawback.Fortunately,the superiorities of the structured DCNs and random DCNs are fully complementary.This motivates us to seek new topologies which can integrate their superiorities together and avoid their weakness.Hence,we propose a family of hybrid topologies,which are compound graphs of given structured and random topologies.(3)Proposing the edge-coloring based routing algorithm for R3.Routing in our hybrid topologies needs a dedicated design method because of the coexistence of random links and structured links.It is clear that the obstacle of routing results from all of random nodes.We thus regularize those random nodes by coloring all of structured links and make routing just like in a totally structured topology.(4)Proposing the novel wireless DCN structure.To realize low-cost and high flexibility for DCNs,we propose a novel wireless DCN structure based on the visible light communication and name it as VLCcube.It extends the design of wireless DCN into three further dimensions: 1)all interrack links are wireless;2)without imposing any infrastructurelevel alteration on wired production data centers;and 3)it should be plug-and-play,without need of additional mechanical or electronic control operations.(5)Designing the routing and scheduling algorithms for the proposed VLCcube.For any pair of ToR switches,wired paths,wireless paths and hybrid paths coexist in VLCcube.The routing algorithms for wired paths and wireless paths can be found in the related literatures.We focus on designing the hybrid routing between racks.To minimize the network congestion,we define a congestion-aware flow scheduling model and design scheduling algorithms for the batched and online traffic patterns.(6)Executing experiments to evaluating our proposals.Comprehensive experiments are conducted to measure the performance of our proposals.The results demonstrate that,compared with Jellyfish,our hybrid structures bring less routing time and cabling cost.Compared with the generalized hypercube,our hybrid structures process better network order and throughput.For VLCcube,the results indicate that VLCcube outperforms Fat-Tree significantly under the existing ECMP flow scheduling scheme,irrespective of the used traffic pattern.Compared to ECMP,our congestion-aware flow scheduling scheme make VLCcube achieve better performance.