ONS 2014: Network Virtualization for Performance Optimization

VMware's Martin Casado defines network virtualization and discusses more use cases.

By Jude Chao | Posted Mar 4, 2014
Page of   |  Back to Page 1
Print ArticleEmail Article
  • Share on Facebook
  • Share on Twitter
  • Share on LinkedIn

Much of the talk and many of the presentations at this year's Open Networking Summit have revolved around SDN. But when it comes to optimizing the new network model that ONS pushes, network virtualization is also part of the story. Talking a mile a minute in an energetic presentation at this morning's SDN for Cloud panel, OpenFlow inventor and VMware chief networking architect Martin Casado delved into some ways VMware has used network virtualization to optimize performance and reduce latency and bottlenecks.

Virtualizaton "tends to capture the imagination," Casado said, even when its utility isn't immediately obvious. In the case of server virtualization, for example, early VMware marketing material tended to claim customers should "buy server virtualization 'cause you get server consolidation, and the reason you want that is so you don't have to buy a lot of servers," he said. It took time for server virtualization's other benefits to emerge. The same holds true for network virtualization. Enterprises tend to adopt it for the faster provisioning times it promises, but the concept offers much more.

Casado defines network virtualization

Before diving into network virtualization's utility, Casado took a moment to define network virtualization. "It's an overloaded term," he explained, adding that "some people think VLANs are network virtualization." To him, they're not.

As with SDN, Casado sticks to a more purist point of view. Network virtualization begins with a physical network, which at minimum requires point-to-point connectivity: "You have to get packets from Point A to Point B," he said. It also requires some edge. The virtual addressing layer then goes on top of the physical, most commonly "by doing tunnel overlays at the edge," he said. An intermediary system manages the mapping of virtual address instances to the physical layer but leaves the addressing decoupled.

"What you want to end up with is things that look like physical networks but have the operational model of virtual machines—you can control all of the state," Casado said.

Runtime network optimization

The decoupling central to the concept of network virtualization will also "decouple the notion of correctness with performance," Casado said. Typically, virtualization overall will require some compromises on performance, he explained, but network virtualization can enable better handling of bottlenecks and bugs.

"There's no reason you can't take a total crap physical network and overlay network virtualization over it," he said. In its initial state, the virtual network won't be ideal, but it will run. Once it's there, however, administrators can make tweaks to improve performance on the fly. To illustrate this, Casado told the crowd about VMware's use of runtime deployment of distributed routing to reduce latency.

"We implemented a fully distributed version of routing so that every little enforcement point was doing a little bit of routing. We updated this in real time on a production network, and if you look at the latency impact, it goes from high to low," he said—allowing the organization to get the most out of their existing physical infrastructure.

Performance optimization via "elephant detection"

Network virtualization can also help organizations maximize the performance of their physical networks by solving the "elephant and mice problem," Casado said. This problem, which he said VMware is working with a number of large organizations to solve, is simple to explain, but difficult to resolve. Within a datacenter, the vast majority of data flows are small (mice), while the majority of packets are in large flows (elephants). A backup or other large workload would be an elephant, he explained, and having an elephant on the network can significantly slow down the numerous latency-sensitive mice.

To date, Casado said, it has been difficult to solve the elephant problem. Many proposed solutions are predictive, which doesn't help when a big, unscheduled workload lumbers onto the network. Network virtualization can help.

"Because network virtualization is normally pushed to the edge in some way, you actually have pretty good visibility in what's going on," he observed. If it's pushed all the way to the vSwitch, meanwhile, "you can detect beforehand the amount of information that's going to go into the network," he added. With that information, the elephants can be split up—turned into mice—and reordered on the other side "without the TCP stack having to go berserk."

VMWare's NSX network virtualization platform hit the market in August of 2013, but Casado has been involved with network virtualization since the Nicira days and has seen "production deployments across the industry for the last three years," long enough to achieve a deeper understanding of network virtualization's benefits. Casado has spoken at length with Enterprise Networking Planet about some of those benefits, such as upgrade and feature velocity and easy snapshotting and modeling of a network's entire state. With today's ONS 2014 panel, he adds a couple more compelling reasons to explore network virtualization technology.

Header photo courtesy of Shutterstock.

ENP editor Jude ChaoJude Chao is executive editor of Enterprise Networking Planet. Follow her on Twitter @judechao.

Comment and Contribute
(Maximum characters: 1200). You have
characters left.
Get the Latest Scoop with Enterprise Networking Planet Newsletter