Building a Flat Network at Data Center Scale

In many ways, today’s global cloud-service providers are setting a new standard for enterprise data centers. Through their ability to deliver high-performance data services at a huge scale and a relatively low cost, they are showing their enterprise counterparts the way to the future.

Lots of this comes down to innovative architectures that leap over some of the barriers that historically restricted enterprise data centers. That’s the way it is with the cloud service providers’ approach to data storage, which is based on software-defined infrastructure (SDI) and storage that is distributed across the data center. This next-generation architecture puts data closer to the processors and eliminates the need for a lot of redundant hardware and software components.

This new approach to storage is ideal for the reality of today’s mobile applications and web services, which tend to be light on “north-south” traffic and heavy on “east-west” traffic. Let me explain what I’m getting at here.

Data center networks have historically been designed to support traffic between an office-resident terminal (and, over time, a PC) and a set of data-center-resident services, such as file, database, and email services. Within the data center, this class of traffic is referred to as “north-south,” because it always flows through a WAN gateway to a server, traversing the core of the data center network.

With the emergence of mobile applications, users increasingly access data-center-resident applications via web services that are rendered on mobile phones, tablets, and laptop computers. Only a small amount of content is delivered to the user’s device. Within the data center, however, many different services are accessed to render the content. This server-to-server communication is referred to as “east-west” traffic, and today we are in the midst of a virtual explosion in this type of traffic.

The east-west traffic pattern shifts the emphasis of the network design to bisection bandwidth, the bandwidth available to support communication between an arbitrary pair of servers attached to the network. With this new architecture, the data center operator can provision an application on any server in the data center, rather than on a single dedicated server.

Better still, when any server can do the work of any other server, the data center operator no longer needs to spend a small fortune to create redundant systems to enable resilient services that meet the performance expectations of end users. The resiliency that used to be gained with redundant hardware-based systems is now achieved with pervasive software, eliminating a huge amount of overprovisioning.

It is important to note that this design is primarily a response to the need to support “application mobility,” or the ability to provision/place an application in the data center anywhere/anytime. From a storage perspective, this translates directly into the need for “application placement” near where the data resides. Such placement schemes are much more about exploiting the “application-to-storage” connectivity model than about the physical construction of the infrastructure.

Google* is a great example of this next-generation approach to the storage landscape. It creates flat networks at data center scale not just with hardware but with hardware combined with software virtual switches on the servers, an end-to-end routing substrate, and centralized software-defined networking (SDN) controllers. For a closer look at what Google is doing in this area, check out the Open Networking Summit 2015 keynote from Amin Vahdat, the technical lead for networking at Google.

Now here is the million-dollar question: How do we enable the rest of the world with this architecture? At Intel, we are working actively deliver the enabling technologies and infrastructure components that will bring cloud-like agility and economics to enterprise storage. These offerings include next-generation Intel® Xeon® processors, new non-volatile memory (NVM) like 3D XPoint technology, and the Storage Performance Development Kit (SPDK).

Here’s the bottom line: Led by the technology leadership of their cloud counterparts, enterprise data centers are on the eve of a transformation that will radically alter the storage landscape to drive higher levels of performance, agility, and efficiency.