Designing and Building Data Center Fabrics

Before starting the Designing and Building Data Center Fabrics online course, let’s try to answer a seemingly simple question “what is a data center fabric?”

What services should your data center fabric provide? What requirements should you consider when designing it? This section lists a number of potential requirements and explains why they might be relevant to your design.

Data center fabrics often rely on technologies you might not find anywhere else in your network. This section describes those technologies and their impact on fabric design.

Traditional data center networks used a 3-tier design that was mostly mandated by hardware limitations, resulting in unequal bandwidth between endpoints based on their locations. In the last few years the networking industry rediscovered the work of Charles Clos (from 1953) and everyone started promoting leaf-and-spine fabrics.

After mastering the basic principles of leaf-and-spine fabrics described in the Introduction section we're moving on to the physical design: how do you build a leaf-and-spine fabric given number of edge ports and oversubscription ratio? What if you need less than 100 ports? What if you need 50.000 ports? What do you do if you have to support low-speed edge interfaces?

We're starting the design part of the webinar with the simplest possible scenario – each leaf switch is a single IP subnet – and focus on routing protocol selection, route summarization, leaf-to-spine link aggregation, and core link addressing.

The second half of this section describes detailed guidelines on using BGP and OSPF as the underlay routing protocols in a leaf-and-spine fabric.

After establishing the baseline in the Layer-3 fabrics with non-redundant server connectivity section we'll add complexity at the fabric edge: redundantly connected servers.

Is it possible to build a pure layer-3 data center fabric that supports redundant server connectivity and IP address mobility? You'll find out in this section.

Another approach to building a pure layer-3 fabric is to extend the fabric routing protocol into the servers and announce servers' loopback IP addresses using BGP.

We're leaving the stable world of L3-only fabrics and entering the realm of large VLANs that most enterprise data centers have to deal with. We'll cover numerous design scenarios, from traditional bridging to routing on layer 2 and MAC-over-IP encapsulation.

Most data center fabrics have to combine elements of large VLANs and routing. In this section we'll explore the various combinations, from traditional routing on spine switches to anycast routing on leaf switches.

Should you stretch a single fabric across multiple sites? Does it make sense to split a large fabric into smaller fabrics (pods)? What could you do to improve the scalability of VXLAN-based EVPN fabrics?

This section contains the design guidelines and technology details you need to answer these questions.

Do you need QoS in data center fabrics? How do you handle load balancing between elephant and mice flows? This section describes some of the answers.

This section describes the differences between SDN, disaggregation and whitebox switching, the major components you'll find in any disaggregated solution (from hardware abstraction layer to control-plane components) and the benefits and drawbacks of several disaggregated network operating systems.

In this section you'll learn how to build very large data center fabrics. The section covers:

• planning and design
• disaggregation of chassis switches into individual components
• challenges and limitations of modern data center switches
• pod architecture
• sample single-SKU designs

As you analyze the vendor solutions, it quickly becomes obvious that every single one of them uses one of few typical control/management plane architectures, from completely decentralized model (sometimes heavily automated) to centralized control- or even data plane.

This section describes most of the architectures you might encounter in real life and documents their benefits and drawbacks.

Buffering is a confusing and sometimes controversial topic when it comes to data center fabrics. In this short section JR Rivers discusses some of the related theory and practicalities as well as gives insight into tools that are available to networking specialists.

In this section you’ll find an overview of data center fabric solutions from major data center switching vendors. The Data Center Fabrics webinar (included with this course) contains even more information.

All webinars in this section are included with your online course. Some of them are part of the core course content, others are included as reference material.