The Role of Networks

A network lives to serve the connectivity requirements of applications and applications. We build networks by designing and implementing data centers. A common trend is that the data center topology is much bigger than a decade ago, with application requirements considerably different from the traditional client-server applications and with deployment speeds in seconds instead of days. This changes how networks and your chosen data center topology are designed and deployed.

The traditional network design was scaled to support more devices by deploying larger switches (and routers). This is the scale-in model of scaling. However, these large switches are expensive and primarily designed to support only a two-way redundancy.

Today, data center topologies are built to scale out. They must satisfy the three main characteristics of increasing server-to-server traffic, scale ( scale on-demand ), and resilience. The following diagram shows a ToR design we discussed at the start of the blog.

The Role of The ToR

Top of rack (ToR) is a term used to describe the architecture of a data center. It is a server architecture in which servers, switches, and other equipment are mounted on the same rack. This allows for the most efficient use of space since the equipment is all within arm’s reach.

ToR is also the most efficient way to manage power and cooling since the equipment is all in the same area. Since all the equipment is close together, ToR also allows faster access times. This architecture can also be utilized in other areas, such as telecommunications, security, and surveillance.

ToR is a great way to maximize efficiency in any data center and is becoming increasingly popular. In contrast to the ToR data center design, the following diagram shows an EoR switch design.

The Role of The EoR

The term end-of-row (EoR) design is derived from a dedicated networking rack or cabinet placed at either end of a row of servers to provide network connectivity to the servers within that row. In EoR network design, each server in the rack has a direct connection with the end-of-row aggregation switch, eliminating the need to connect servers directly with the in-rack switch.

Racks are usually arranged to form a row; a cabinet or rack is positioned at the end of this row. This rack has a row aggregation switch, which provides network connectivity to servers mounted in individual racks. This switch, a modular chassis-based platform, sometimes supports hundreds of server connections. However, a large amount of cabling is required to support this architecture.

A ToR configuration requires one switch per rack, resulting in higher power consumption and operational costs. Moreover, unused ports are often more significant in this scenario than with an EoR arrangement.

On the other hand, ToR’s cabling requirements are much lower than those of EoR, and faults are primarily isolated to a particular rack, thus improving the data center’s fault tolerance.

If fault tolerance is the ultimate goal, ToR is the better choice, but EoR configuration is better if an organization wants to save on operational costs. The following table lists the differences between a ToR and an EoR data center design.

Data Center Topology Types:

Fabric extenders – FEX

Cisco has introduced the concept of Fabric Extenders, which are not Ethernet switches but remote line cards of a virtualized modular chassis ( parent switch ). This allows scalable topologies previously impossible with traditional Ethernet switches in the access layer.

You should relate an FEX device like a remote line card attached to a parent switch. All the configuration is done on the parent switch, yet physically, the fabric extender could be in a different location. The mapping between the parent switch and the FEX ( fabric extender ) is done via a special VN-Link.

The following diagram shows an example of a FEX in a standard data center network topology. More specifically, we are looking at the Nexus 2000 FEX Series. Cisco Nexus 2000 Series Fabric Extenders (FEX) are based on the standard IEEE 802.1BR. They deliver fabric extensibility with a single point of management.

Different types of Fex solution

FEXs come with various connectivity solutions, including 100 Megabit Ethernet, 1 Gigabit Ethernet, 10 Gigabit Ethernet ( copper and fiber ), and 40 Gigabit Ethernet. They can be synchronized with the following parent switch models: Nexus 5000, Nexus 6000, Nexus 7000, Nexus 9000, and Cisco UCS Fabric Interconnect.

In addition, because of FEX’s simplicity, they have very low latency ( as low as 500 nanoseconds ) compared to traditional Ethernet switches.

Some network switches can be connected to others and operate as a single unit. These configurations are called “stacks” and are helpful for quickly increasing the capacity of a network. A stack is a network solution composed of two or more stackable switches. Switches that are part of a stack behave as one single device.

Traditional switches like the 3750s still stand in the data center network topology access layer and can be used with stacking technology, combining two physical switches into one logical switch.

This stacking technology allows you to build a highly resilient switching system, one switch at a time. If you are looking at a standard access layer switch like the 3750s, consider the next-generation Catalyst 3850 series.

The 3850 supports BYOD/mobility and offers various performance and security enhancements compared to previous models. However, stacking has a drawback: You can only stack several switches. So, if you want more throughout, you should aim for a different design type.

Data Center Design: Layer 2 and Layer 3 Solutions

Traditional views of data center design

Depending on the data center network topology deployed, packet forwarding at the access layer can be either Layer 2 or Layer 3. A Layer 3 approach would involve additional management and configuring IP addresses on hosts in a hierarchical fashion that matches the switch’s assigned IP address.

An alternative approach is to use Layer 2, which has less overhead as Layer 2 MAC addresses do not need specific configuration. However, it has drawbacks with scalability and poor performance.

Generally, access switches focus on communicating servers in the same IP subnet, allowing any type of traffic – unicast, multicast, or broadcast. You can, however, have filtering devices such as a Virtual Security Gateway ( VSG ) to permit traffic between servers, but that is generally reserved for inter-POD ( Platform Optimized Design ) traffic.

Leaf and Spine With Layer 3

We use a leaf and spine data center design with Layer 3 everywhere and overlay networking. This modern, robust architecture provides a high-performance, highly available network. With this architecture, data center networks are composed of leaf switches that connect to one or more spine switches.

The leaf switches are connected to end devices such as servers, storage devices, and other networking equipment. The spine switches, meanwhile, act as the network’s backbone, connecting the multiple leaf switches.

The leaf and spine architecture provides several advantages over traditional data center networks. It allows for greater scalability, as additional leaf switches can be easily added to the network. It also offers better fault tolerance, as the network can operate even if one of the spine switches fails.

Furthermore, it enables faster traffic flows, as the spine switches to route traffic between the leaf switches faster than a traditional flat network.

Data Center Traffic Flow

Datacenter topologies can have North-South or East-to-West traffic. North-south ( up / down ) corresponds to traffic between the servers and the external world ( outside the data center ). East-to-west corresponds to internal server communication, i.e., traffic does not leave the data center.

Therefore, determining the type of traffic upfront is essential as it influences the type of topology used in the data center.

For example, you may have a pair of ISCSI switches, and all traffic is internal between the servers. In this case, you would need high-bandwidth inter-switch links. Usually, an ether channel supports all the cross-server talk; the only north-to-south traffic would be management traffic.

In another part of the data center, you may have data server farm switches with only HSRP heartbeat traffic across the inter-switch links and large bundled uplinks for a high volume of north-to-south traffic. Depending on the type of application, which can be either outward-facing or internal, computation will influence the type of traffic that will be dominant. 

Virtual Machine and Containers.

This drive was from virtualization, virtual machines, and container technologies regarding east-west traffic. Many are moving to a leaf and spine data center design if they have a lot of east-to-west traffic and want better performance.

Network Virtualization and VXLAN

Network virtualization and the ability of a physical server to host many VMs and move those VMs are also used extensively in data centers, either for workload distribution or business continuity. This will also affect the design you have at the access layer.

For example, in a Layer 3 fabric, migrating a VM across that boundary changes its IP address, resulting in a reset of the TCP sessions because, unlike SCTP, TCP does not support dynamic address configuration. In a Layer 2 fabric, migrating a VM incurs ARP overhead and requires forwarding on millions of flat MAC addresses, which leads to MAC scalability and poor performance problems.

VXLAN: stability over Layer 3 core

Network virtualization plays a vital role in the data center. Technologies like VXLAN attempt to move the control plane from the core to the edge and stabilize the core so that it only has a handful of addresses for each ToR switch. The following diagram shows the ACI networks with VXLAN as the overlay that operates over a spine leaf architecture.

Layer 2 and 3 traffic is mapped to VXLAN VNIs that run over a Layer 3 core. The Bridge Domain is for layer 2, and the VRF is for layer 3 traffic. Now, we have the separation of layer 2 and 3 traffic based on the VNI in the VXLAN header.  

One of the first notable differences between VXLAN and VLAN was scale. VLAN has a 12-bit identifier called VID, while VXLAN has a 24-bit identifier called a VID network identifier. This means that with VLAN, you can create only 4094 networks over ethernet, while with VXLAN, you can create up to 16 million.

Whether you can build layer 2 or layer 3 in the access and use VXLAN or some other overlay to stabilize the core, it would help if you modularized the data center. The first step is to build each POD or rack as a complete unit. Each POD will be able to perform all its functions within that POD.

• A key point: A POD data center design

POD is a design methodology that aims to simplify, speed deployment, optimize resource utilization, and drive the interoperability of three or more data center components: server, storage, and networks.

• A POD example: Data center modularity

For example, one POD might be a specific human resources system. The second is modularity based on the type of resources offered. For example, a storage pod or bare metal compute may be housed in separate pods.

These two modularization types allow designers to easily control inter-POD traffic with predefined policies. Operators can also upgrade PODs and a specific type of service at once without affecting other PODs.

However, this type of segmentation does not address the data center’s scale requirements. Even when we have adequately modularized the data center into specific portions, the MAC table sizes on each switch still increase exponentially as the data center grows.

Current and Future Design Factors

New technologies with scalable control planes must be introduced for a cloud-enabled data center, and these new control planes should offer the following:

Several design factors need to be considered when designing a data center. First, what is the growth rate for servers, switch ports, and data center customers? This prevents part of the network topology from becoming a bottleneck or linking congested.

**Application bandwidth demand?**

This demand is usually translated into oversubscription. In data center networking, oversubscription refers to how much bandwidth switches are offered to downstream devices at each layer.

Oversubscription is expected in a data center design. Limiting oversubscription to the ToR and edge of the network offers a single place to start when performance problems occur.

A data center with no oversubscription ratio will be costly, especially with a low latency network design. So, it’s best to determine what oversubscription ratio your applications support and work best. Optimizing your switch buffers to improve performance is recommended before you decide on a 1:1 oversubscription rate.

**Ethernet 6-byte MAC addressing is flat**

Ethernet forms the basis of data center networking in tandem with IP. Since its inception 40 years ago, Ethernet frames have been transmitted over various physical media, even barbed wire. Ethernet 6-byte MAC addressing is flat; the manufacturer typically assigns the address without considering its location.

Ethernet-switched networks do not have explicit routing protocols to ensure readability about the flat addresses of the server’s NICs. Instead, flooding and address learning are used to create forwarding table entries.

**IP addressing is a hierarchy**

On the other hand, IP addressing is a hierarchy, meaning that its address is assigned by the network operator based on its location in the network. A hierarchy address space advantage is that forwarding tables can be aggregated. If summarization or other routing techniques are employed, changes in one side of the network will not necessarily affect other areas.

This makes IP-routed networks more scalable than Ethernet-switched networks. IP-routed networks also offer ECMP techniques that enable networks to use parallel links between nodes without spanning tree disabling one of those links. The ECMP method hashes packet headers before selecting a bundled link to avoid out-of-sequence packets within individual flows. 

Equal Cost Load Balancing

Equal-cost load balancing is a method for distributing network traffic among multiple paths of equal cost. It provides redundancy and increases throughput. Sending traffic over numerous paths avoids congestion on any single link. In addition, the load is equally distributed across the paths, meaning that each path carries roughly the same total traffic.

This allows for using multiple paths at a lower cost, providing an efficient way to increase throughput.

The idea behind equal-cost load balancing is to use multiple paths of equal cost to balance the load on each path. The algorithm considers the number of paths, each path’s weight, and each path’s capacity. It also considers the number of packets that must be sent and the delay allowed for each packet.

Considering these factors, it can calculate the best way to distribute the load among the paths.

Equal-cost load balancing can be implemented using various methods. One method is to use a Link Aggregation Protocol (LACP), which allows the network to use multiple links and distribute the traffic among the links in a balanced way.

• A keynote: Data center topologies. The move to VXLAN.

Given the above considerations, a solution encompassing the benefits of L2’s plug-and-play flat addressing and IP scalability is needed. Location-Identifier Split Protocol ( LISP ) has a set of solutions that use hierarchical addresses as locators in the core and flat addresses as identifiers in the edges. However, not much is seen in its deployment these days.

Equivalent approaches such as THRILL and Cisco FabricPath create massive scalable L2 multipath networks with equidistant endpoints. Tunneling is also being used to extend down to the server and access layer to overcome the 4K limitation with traditional VLANs. What is VXLAN? Tunneling with VXLAN is now the standard design in most data center topologies with leaf-spine designs. The following video provides VXLAN guidance.

Data Center Network Topology

Leaf and spine data center topology types

This is commonly seen in a leaf and spine design. For example, in a leaf-spine fabric, We have a Layer 3 IP fabric that supports equal-cost multi-path (ECMP) routing between any two endpoints in the network. Then, on top of the Layer 3 fabric is an overlay protocol, commonly VXLAN.

A spine-leaf architecture consists of a data center network topology with two switching layers: a spine and a leaf. The leaf layer comprises access switches that aggregate traffic from endpoints such as servers and connect directly to the spine or network core.

Spine switches interconnect all leaf switches in a full-mesh topology. The leaf switches do not directly connect. The Cisco ACI is a data center topology that utilizes the leaf and spine.

The ACI network’s physical topology is a leaf and spine, while the logical topology is formed with VXLAN. From a protocol side point, VXLAN is the overlay network, and the BGP and IS-IS provide the Layer 3 routing, the underlay network that allows the overlay network to function.

As a result, the nonblocking architecture performs much better than the traditional data center design based on access, distribution, and core designs.

**Closing Points: Data Center Topologies**

A data center topology refers to the physical layout and interconnection of network devices within a data center. It determines how servers, switches, routers, and other networking equipment are connected, ensuring efficient and reliable data transmission. Topologies are based on scalability, fault tolerance, performance, and cost.

• Hierarchical Data Center Topology:

The hierarchical or tree topology is one of the most commonly used data center topologies. This design consists of multiple core, distribution, and access layers. The core layer connects all the distribution layers, while the distribution layer connects to the access layer. This structure enables better management, scalability, and fault tolerance by segregating traffic and minimizing network congestion.

• Mesh Data Center Topology:

Every network device is interlinked in a mesh topology, forming a fully connected network with multiple paths for data transmission. This redundancy ensures high availability and fault tolerance. However, this topology can be cost-prohibitive and complex, especially in large-scale data centers.

• Leaf-Spine Data Center Topology:

The leaf-spine topology is gaining popularity due to its scalability and simplicity. It consists of interconnected leaf switches at the access layer and spine switches at the core layer. This design allows for non-blocking, low-latency communication between any leaf switch and spine switch, making it suitable for modern data center requirements.

• Full-Mesh Data Center Topology:

As the name suggests, the full-mesh topology connects every network device to every other device, creating an extensive web of connections. This topology offers maximum redundancy and fault tolerance. However, it can be expensive to implement and maintain, making it more suitable for critical applications with stringent uptime requirements.