table of Contents
Article Directory
The foregoing list
" Problem with the evolution of the data center network architecture - traditional routing and switching technology three-tier network architecture ,"
" the problem of data center network architecture and evolution - network virtualization "
" problems with the evolution of the data center network architecture - CLOS network and Fat-Tree , Spine-Leaf architecture "
Data Center Network Architecture Evolution Review
The traditional three-tier data center (Three-tier Architecture), according to the physical location of the server into a different POD. This architecture consists of core switches, and access switches composition aggregation. Spanning Tree Protocol (Spanning Tree Protocol, STP) between aggregation and access switches to ensure that part of the network Layer (L2) without loops. Three-tier architecture, the server within each POD belong to the same VLAN, so the server without modifying the IP address and gateway can migrate to any location within the POD. Broadcast take the POD Layer, Layer walking route between the POD, the lateral flow path will be long (the access aggregation → → → heart → access aggregation) caused by high latency. STP has many advantages: simple, plug and play (plug-and-play, P & P), with minimal configuration. However, STP can not use parallel forwarding path (Parallel Forwarding Path), so it will always be disabled redundant paths in the VLAN, which means that when a fault has not occurred, 50% of the maximum link bandwidth will be Block (blocked) away.
In 2010, Cisco made vPC (Virtual Port Channel, virtual link aggregation) technology to solve STP restrictions. vPC liberated STP port is disabled, the switch providing access to the Active-Active uplink between the aggregation switch, make full use of available bandwidth. When using the vPC technology, STP as an alternate mechanism (Fail-safe Mechanism). Development of a virtual machine technology, and other containers, so that the deployment of distributed applications more and more, resulting in a lateral flow inside the data center (East-West-Traffic) significantly improved. The flow needs to be efficiently treated, and also to ensure a low, predictable delay. However, the vPC uplink only provide two parallel, so three-tier data center architecture bandwidth has become a bottleneck.
With the development of virtualization technology, the original three-tier network architecture of the data center floor POD-based design, has been unable to meet the virtualization, cloud computing needs of large-scale two-story extension, for example: vMotion technologies such as Live Migration requires more L2 large as possible. In so doing Layer POD in the form of the isolated computing, network and storage resources, it should be pooled (Pooled, and then spawned a large floor layer from the domain access to the core layer (Larger L2 Domain). Large Layer 2 zones, namely the entire data center network is a broadcast domain L2 (L2 Segment is extended to all POD). in this way, data center managers can create a centralized, more flexible, able-demand resource pool , server / virtual machines can be created in any location, migration, without the need to modify the IP address or the default gateway.
Disadvantage of large Layer 2 network is BUM (Broadcase, Unknown Unicast, Multicast) will be spread throughout the data center, which ultimately limits the size of the network. Because of the large size of the network to a certain extent, BUM storm will seriously affect the normal network communications. This is another new data center network development issues raised during the network service providers, and service providers have to surrender the respondents is Overlay Network.
Overlay Network
现代虚拟化数据中心的网络要加速应用部署和支持 DevOps，必须满足特定的前提条件。例如：需要支持扩展转发表、扩展网段、L2 Segment Extension、虚拟设备漂移（Mobility）、转发路径优化、共享物理基础设施上的网络虚拟化和多租户等等。Overlay 网络可以解决以上提到的问题，因此各厂家在今年来陆续推出了多种新的帧封装格式（Encapsulation Frame Format）：
- VXLAN
- GRE
- Geneva
- NVGRE: Network Virtualization Using Generic Routing Encapsulation
- TRILL: Transparent Interconnection of Lots of Links
- LISP: Location/Identifier Separation Protocol
Overlay 网络是指在现有的网络（Underlay Network）的基础上再构建的一个虚拟网络。所谓的 Underlay 网络，就是之前的交换机所在的网络，只要是 IP 网络就行。而新构建的 Overlay 网络，用来作为服务器通讯的网络。Overlay 网络是一个 L3 in L2 网络。也就是说，只要 L3 网络能覆盖的地方，那 Overlay 的 L2 网络也能覆盖。
简而言之，Overlay 网络就是是在共享底层网络（Underlay Network）的节点之间互连形成的虚拟网络，这使得在不修改底层网络的情况下，可以部署对网络拓扑有特定要求的应用。
Overlay 网络的优势
优化的设备功能：Overlay 网络使得可以根据设备在网络中的位置不同而对设备进行分类（和定制）。Edge 或 Leaf 设备可以根据终端状态信息和规模优化它的功能和相关的协议；Core 或 Spine 设备可以根据链路状态优化它的功能和协议，以及针对快速收敛进行优化。
Fabric 的扩展性和灵活性：Overlay 网络使得可以在 Overlay 边界设备上进行网络的扩展。当在 Fabric 边界使用 Overlay 时，Spine 或 Core 设备就无需向自己的转发表中添加终端主机的信息（e.g. 如果在宿主机内进行 Overlay 的封装和解封装，那 Overlay 边界就是在宿主机内部）。
可重叠的 IP 地址：数据中心中使用的大部分 Overlay 技术都支持虚拟网络 ID（VID，），用来唯 一地对每个私有网络进行范围限定和识别（scope and identify）。这种限定使得不同租 户的 MAC 和 IP 地址可以重叠（overlapping）。overlay 的封装使得租户地址空间和 underlay 地址空间的管理分开。
基于 VxLAN Overlay 的 Spine-Leaf 网络架构
基于 VxLAN 的 Overlay 网络在 L3 IP Underlay 网络之上构建了一层 L2 Overlay 网络，通过 VTEP 隧道机制传输 L2 包。例如下图中，原有的交换机网络不变，服务器之间通过 Overlay 网络实现了跨 Leaf 交换机的 L2 网络。这样，在 Overlay 网络中，服务器可以任意部署，而不用考虑现有网络的架构。
- 单纯的 Spine-Leaf
- 基于 VxLAN Overlay 的 Spine-Leaf
A complete, based on FIG. Spine-Leaf VxLAN Overlay network architecture as follows. This network architecture is called VxLAN Fabric, usually implemented in two ways: one is based on Flood-Learn mode, similar to the conventional L2 network; the other is based on the MP-BGP EVPN as the control layer. In this architecture, the VM does not know of the existence VxLAN Overlay, VM just issued to the Ethernet Frame. After VTEP Leaf switch to be on (the Overlay VxLAN, need to be integrated on a Leaf VTEP switch) receives the VM Ethernet Frame, will automatically VxLAN encapsulates the data packet (essentially a UDP packet), then the original Spine L3 -Leaf for transmission on the network Underlay. Spine-Leaf-based network architecture VxLAN Overlay is through this way to break the limit of large Layer 2 network, the upcoming L2 broadcast domain restricted on Underlay (Leaf node), in turn POD extended to the whole of the Overlay .
About VxLAN technology is implemented in principle " Networking basic terms / concepts have been recorded," a text, not repeat them here.
Reference article
https://arthurchiao.github.io/blog/spine-leaf-design-zh/
https://zhuanlan.zhihu.com/p/30119950