Smart Internet

The increase in the scale of data centers directly leads to a leapfrog upgrade in network complexity. Under the tide of new infrastructure, the demand of enterprises and the society for various types of digital infrastructure is unabated. To allow data and business to flow smoothly in the data center, an efficient network that matches the size of the data center is the key. The first step in building an efficient network is to choose the right network architecture.

Three data center scenarios, three networking models

Along with the deepening of digital transformation, "scenario" has become a buzzword in the past two years. While computing and storage devices are still racking their brains over the needs and design of different scenarios, the scenarioization of the network has long been a consensus in the industry.

Just as people will build different levels of highways to face different access needs, engineers have long planned optimal networking solutions for different sizes and applications of data centers.

Scenario 1: Small and Medium Data Center

The two-tier Clos architecture is the earliest and most common network architecture and is still the preferred choice for many customers in many industries. For small to medium sized data centers, which are generally limited in size, this simple architecture still allows us to build reliable networks with ever-increasing switch specifications. This is the main reason why the two-tier Clos architecture is so popular in small and medium-sized data centers.

Typical Two-tier Clos Architecture: H3C AD-DC Application Driven Data Center Solution

In the two-level Clos architecture, there are only two roles for the whole network equipment. The advantages of this architecture are short data forwarding paths, one-hop reachability across the Leaf, and strong consistency in paths and delays. The unified access method also brings great convenience to on-line deployment and horizontal expansion, such as the deployment of BGP protocol, policy control, daily maintenance and problem troubleshooting.

The two-tier Clos architecture requires high performance and reliability of Spine switches, which are generally used in data center boxed core switch products. Boxed core switches have independent control planes, forwarding planes, and support systems, and are designed for redundancy, which makes the entire system far more reliable than boxed switches.

Performance Evolution of Data Center Boxed Core Switches from 576*10G to 576*400G

The two-tier Clos architecture is more mature in adapting to commercial SDN controller solutions. Combined with SDN controllers, it can quickly build EVPN-based network overlay solutions, reduce the difficulty of deploying east-west and north-south service chains, and satisfy the needs of the network in cloud scenarios for the linkage of full-format computing resources, such as VMs, bare metals, containers, and so on.

In addition, the architecture is equally suitable for aggregation and edge rooms deployed by large enterprises in various locations to build edge computing networks, relieve backbone network pressure and reduce access latency.

Scenario 2: Medium to Large Data Center

The size of servers supported by the two-tier Clos architecture is generally less than 20,000, and the introduction of the three-tier Clos architecture solves the bottleneck of the two-tier Clos architecture in terms of network size. Three-tier Clos architecture adds a level of aggregation switch (Pod Spine) in the middle of the two-tier Clos architecture, which consists of a group of Pod Spine switches and all the Leaf switches connected to them to form a Pod, and interconnects multiple Pods to form the whole network through the Spine layer switches. Increasing the number of Pods can realize the horizontal expansion of the network, which greatly improves the network scalability. At the same time, the three-tier Clos architecture is more flexible in terms of adapting to multiple service requirements and providing differentiated services by deploying services on a pod basis.

Example of a three-tier Clos architecture

In addition, due to the introduction of Pod Spine, a high-density aggregation switch, the box core switch at the Spine layer breaks through the single-digit limit and can be deployed in dozens of units. The total number of ports provided by the box core switch at the Spine layer can be used for connecting to dozens of Pods, and the entire network can support the scale of more than 100,000 servers.

Performance evolution of high density aggregation switches from 32*100G to 64*400G

In addition, the convergence ratio of each Pod can be flexibly defined by adjusting the ratio of upstream and downstream ports of Pod Spine switches within the Pod, which also helps to reduce costs and avoid unnecessary waste while meeting different business requirements.

Scenario 3: Large and Very Large Data Centers

the InternetEnterprises are always refreshing the term "massive", both at the business model level and at the infrastructure level. Every minute hundreds of millions or even billions of business on the scale of the data center puts forward higher requirements, and this has led to another evolution of the network structure. Multi-plane networking architecture based on boxed devices is the latest architecture adopted by current head Internet companies for the formation of large-scale and ultra-large-scale data center networks.

Example of multiplane networking based on boxed devices

Unlike the three-tier Clos architecture where each Pod Spine needs to be fully interconnected with all Spine layer switches, the Spine layer switches in the new architecture are divided into multiple groups (the number of groups is the same as the number of Pod Spine switches in each Pod), and the Spine switches in each group can form an independent plane, and the Pod Spine switches in each Pod only need to be fully interconnected with the Spine switches in the corresponding plane. Spine switches in each Pod can form an independent plane, and the Pod Spine switches in each Pod only need to be fully interconnected with the Spine switches in the corresponding plane.

In this way, the entire Spine layer can be connected to more Pods, building a hyperscale network that supports hundreds of thousands of class servers. And, as the performance of the box switch increases, the architecture can continue to increase the capacity space.

At the same time, since Spine and Pod Spine use the same equipment, the entire network has a high degree of consistency in terms of functionality and forwarding latency. This creates a huge advantage for service deployment and network tuning. Moreover, the entire network can be synchronized with the evolution from 100G to 200G, 400G and subsequent higher-speed networks.

Future Data Center Networks

01、Visualized network management

From the above three different sizes of data center networking solutions we can find that no matter what kind of architecture, management is a core issue that cannot be ignored. Then, the question arises: how to achieve more efficient network management at a lower cost?

Seeing the traffic allows us to better manage the roads; visualization of the network is also a prerequisite for efficient management.

In practical applications, network visualization technology can not only complete end-to-end traffic monitoring, risk warning, and assist in troubleshooting; it can also achieve the optimization design of data center network architecture through data accumulation and analysis.

02, smart network card will become the new end of the network

Switch + SmartNIC + UniServer Server End-to-End Solution

In the future, smart NICs will be an important part of DCN networks. Smart NICs with programmable capabilities will release CPU resources and realize high-performance forwarding while also possessing functions such as tunnel encapsulation/decapsulation, virtual switching, encryption/decryption, RDMA, etc. With the increase in business scenarios and demands, more and more data plane functions will be completed by smart NICs, breaking the limitations realized by each of them based on server or switch, which is expected to achieve a perfect balance of performance and functionality. This breaks the limitations of server-based or switch-based implementations and is expected to achieve a perfect balance of performance, functionality and flexibility.

In the digital age, cyber progress never stops

China has built nearly 5 million kilometers of various types of roads in 70 years, and this has directly contributed to the economic takeoff of the society. Similarly, in the development of data centers, the pursuit of efficient networks by enterprises is never-ending. As a leader in digital solutions, the group under Ziguang, which has risen to the top of the networking field, also has a deep insight into the relationship between business value, data centers and networks. High-efficiency data center network connects the entire chain from data to business to value in the digital era.

The Group will explore the new frontier of network technology with thousands of industries and pave a spacious road of data and value for the new infrastructure and the entire digital era.