VXLAN – One Tunnel to Rule Them All

I often feel like half my job is just keeping up with the new advancements and offerings that constantly churn through my inbox. I think the blogosphere likes to call it… “Disruption.”  Some trends don’t make it much further than the marketing department’s brainstorming session, but some actually stick and begin to be widely accepted as practice.

From my perspective, the trends that don’t persist are ones that don’t solve a “real” problem – sure, there might be marginal benefits to the technology, but not enough to change buying behaviors. The trends that stick are the ones that actually solve a big underlying issue, opening up greater efficiency and productivity. Server virtualization is a great example of a trend that “stuck.” When virtualization was first getting started, there were a lot of naysayers and many that didn’t like how it was changing the datacenter, but it made a very compelling argument for itself with the resiliency, power savings, and hardware savings that it offered. As a result, server virtualization is more or less standard these days, with software-defined storage following it closely.

There is one element of IT that is heavily lagging when it comes to modernization though – networking. If you look at some underlying principles behind network engineering, it makes sense. Networking at the core is a distributed communications system that uses very established protocols. No one entity “owns” the web, nor are updates rolled out across the world simultaneously. Tinkering with a protocol to “improve things” is more likely to cut you off from the rest of the world than help modernize your infrastructure. And in case of misconfiguration, a network outage is HIGHLY visible – if the network is down, so is everything else!

So, for those reasons, the network industry has been historically resistant to change. Configurations are still often done via the CLI. There is a steep tribal knowledge barrier to entry, and because networks can be taken down by forgetting a single keyword, any changes made to the network are tightly regulated. Networking follows the “if it ain’t broke, don’t fix it” mantra and that mantra has served it pretty well… for the most part.

I know I’m venturing into generalization, but I think networks are happiest when they are architected, configured… and then left alone. Nobody wants to be constantly logging in to the datacenter network infrastructure and making changes due to the massive blast radius when human error inevitably strikes.  Human error is one of the leading causes for network outages, after all.

So, this mindset causes several issues when working with large “webscale” datacenters. Some of the key advantages that a virtualized datacenter brings is resilience, efficiency and optimization. If VMs start to run into problems at the host level, a well-designed VMware environment will work around it seamlessly – via vMotion, HA, DRS, and other key technologies – and the users will never realize what’s happening behind the scenes. They just know that everything works and will continue to work. The VMware datacenter can be constantly refreshing, updating, optimizing, shuffling, and so on – all in the name of user experience and efficiency!

The problem is – how does this interact with a network that is designed to be static? Often the network is statically configured in a ‘set-it-and-forget-it’ style. VMware doesn’t natively have a great way to tell OSPF that a VM in subnet A has moved to the other side of the datacenter because it will run better over there. So, instead the virtualization engineers asked the network engineers to extend L2 everywhere so communication wouldn’t be broken… and the network engineers wailed and gnashed their teeth, but it was all for naught as the project had already been approved and the network engineers hadn’t been invited to the initial planning meetings. L2 extension does allow VMware to be as dynamic as it pleases, because RARP allows VMs to move around on the network and they don’t have to be re-IPed each time they move (which would cause all kinds of complexity with user access as well)… but it comes at a cost.

Those of you familiar with networking no doubt recognize some of the pain that is inherent with L2 communications. It’s fast, sure – but in its raw state, volatile. A L2 segment can often rightly be considered a failure domain. Broadcast storms cause headaches, MAC tables have hardware limitations, some flavor of STP is a necessity, and so on. L2 is only meant to have a single path for data, which throws conventional redundancy options out the window. L3 by comparison is stable and resilient. During my slow steady march to the dark side of datacenter networking I had to learn all these complex ways that companies were trying to solve the L2 datacenter problem in network hardware – SPB, TRILL, Fabricpath, VPLS – the hits kept on coming. And frankly, while I know that these are running happily in many production environments, they all seemed very complex to me and they required some expensive hardware and licenses just to function. (And to make matters worse, the one that made the most sense to me, TRILL, developed by Radia Perlman who introduced the original Spanning Tree, was just announced as dead by some rather prominent individuals in the networking space due to vendor ASIC limitations. That’s a bummer.).

I know that following my thought pattern is a bit of a rabbit trail here, but recall what I mentioned at the start of this blog – in order for a technological trend to catch, it has to solve a real problem. And hopefully I’ve illustrated that there is a real problem when pairing a dynamic environment with an environment that resists change.

Enter VXLAN, a newer protocol that is gaining a lot of steam. At the core, VXLAN is another tunneling protocol that can encapsulate frames inside of packets. It orchestrates this through VXLAN Tunnel End Points (VTEPs), which act as entry points into the datacenter fabric and construct and tear down tunnels between each other to shuttle traffic around. The concept of tunneling isn’t new and VXLAN is not vendor proprietary; it was originally created in collaboration by Arista, Cisco and VMware and it is being utilized by many vendors today. It allows the creation of an “Overlay” network that can be defined in software and runs over the “Underlay” network that is created via physical hardware.

The “Underlay” network can be constructed with a L3 design in a rock solid, resilient configuration. Network engineers only have to ensure that there is basic IP connectivity across the datacenter, and that the MTU is large enough to accommodate the VXLAN header that is added to the frame. Once the Underlay is in place, it will not often have to change.

The “Overlay” network is where the real magic happens. Using VXLAN, software can rapidly construct tunnels that ride on top of the “Underlay” network, shaping the network as required by applications. If a VM needs to think that it has L2 adjacency to a VM on the other side of the datacenter, the VTEPs can construct a L2 tunnel between the two VMs. No hardware changes are required and again, the physical network hardware only needs to give basic IP connectivity between the two hosts. Even better, because the VTEPs in software keep track of MAC address tables and L2 connectivity, your hardware only needs to remember where the VTEPs themselves are and how to move traffic between the VTEPs. Everything else is out of sight and out of mind of your physical infrastructure. The real weight of the MAC address table is now in software, rather than inflating the cost of our hardware.

So, as I mentioned, there are several solutions out there that utilize VXLAN – Arista, Cisco’s ACI, VMware NSX, and others. We at Edge are particularly excited to be working with the VMware NSX platform (I believe almost our entire engineering and technical team hold the VCP-NV badge of honor and yours truly is aggressively pursuing it). VMware has been virtualizing servers and abstracting storage for years now; it only makes sense that they would also tackle the last piece of the trifecta – the network!

VMware networking in its raw state has some limitations. First, the standard vSS/vDS is not able to route traffic – it can only switch traffic. This results in some less than ideal traffic patterns, as traffic is hairpinned on the physical network to move between VMs in a segmented multi-tier application. In addition, the physical network has to be configured to support changes in the virtualized network environment… and if network change control is in place (as it should be), this can take a long time!

VMware NSX solves these issues by bringing routing, security, edge services, load balancing, NAT and more down into the virtualized environment. Virtualization administrators can now spin up all the network services that they need within the virtual environment. By bringing L3 intelligence all the way down into the host, the hairpinning problem is solved. By enforcing network traffic policies at the vnic level, even L2 East/West traffic is now secured.

As a result, the physical network is more or less abstracted with VMware NSX. If VMs need to be on the same subnet to operate but they get moved to separate hosts, VTEPs on each host construct a VXLAN tunnel between the two hosts to allow the traffic to pass through as if they were still L2 adjacent. These tunnels are architected by the VMware NSX platform in software, allowing them to fulfill the promise of a Software Defined Data Center.

VMware NSX consists of several key components:

  • NSX Manager
  • NSX Controllers
  • NSX vSwitch
  • NSX Edge Services Gateway

The NSX Manager integrates with vCenter Server to coordinate management across the VMware environment. It’s important to note that if the NSX Manager goes down or otherwise becomes unavailable, NSX will continue to function.

The NSX Controllers share the burden of the control plane and coordinate network functions across the vSphere, ensuring that changes are kept in sync.

The NSX vSwitch is where L2, L3 and firewall decisions are made – within the host! This piece of NSX is responsible for a great deal of what gives the platform it’s “teeth,” so to speak.

Finally, the NSX Edge Services Gateway provides additional services that are not included in the vSwitch – services like IPsec VPN, NAT, load balancing, DHCP, DNS relay, and more. Edge Services Gateways can also be deployed as Perimeter Edges, which bridge the gap between the physical network and your virtual environment. Using dynamic routing protocols like BGP, OSPF or IS-IS they can coordinate changes between the two environments and advertise any changes you make in the NSX environment.

All these components work together to create a software defined data center. Finally the last piece of the puzzle, networking, has been virtualized and can collaborate properly with the VMware environment. It’s starting to take off, too – if you watched VMware’s earning report a few weeks ago, you’ll know that NSX is on track to be a billion dollar product line by the end of 2017. That’s probably artificially boosted by the steep price tag that NSX carries, but hey, still impressive.

There are many other benefits that NSX offers the datacenter that I haven’t covered in this blogpost – service-insertion for traffic filtering, microsegmentation, and so on – and there are many different ways that it can be implemented. Really, I’ve only scratched the surface. But I’ve rambled on more than long enough at this point. If you’d like to learn more about VMware NSX and start going into specifics about how it will interact with your environment – drop me a line and I’d be happy to compare notes. I have a bunch of big books filled with all the specifics, fine print, addendum and nerd knobs.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s