Category: Data Center

Facebook published their next-generation data center architecture a few weeks ago, resulting in the expected “revolutionary approach to data center fabrics” echoes from the industry press and blogosphere.

In reality, they did a great engineering job using an interesting twist on pretty traditional multi-stage leaf-and-spine (or folded Clos) architecture.

Simonp made a perfectly valid point in a comment to my latest OVS blog post:

Obviously the page you're referring to is a quick-and-dirty benchmark. If you wanted the optimal numbers, you would have to tune quite a few parameters just like for hardware benchmarks (sysctl kernel parameters, Jumbo frames, ...).

While he’s absolutely right, this is not the performance data a typical user should be looking for.

Overlay virtual networks are one of my favorite topics – it seems I wrote over a hundred blog posts describing various aspects of this emerging (or is it reinvented) technology since Cisco launched VXLAN in 2011.

During the summer of 2014 I organized my blog posts on overlay networks and SDDC into a digital book. I want to make this information as useful and as widely distributed as possible – for a limited time you can download the PDF free of charge.

I carried out an interesting quiz during one of my Interop workshop:

• How many use Linux-based servers? Almost everyone raised their hands;
• How many use Apache or Tomcat web servers? Yet again, almost everyone.
• How many run applications written in PHP, Python, Ruby…? Same crowd (probably even a bit more).
• How many use Nginx, Squid or HAProxy for load balancing? Very few.

Is there a rational explanation for this seemingly nonsensical result?

It doesn’t make sense to build a new data center network to support legacy bare-metal server infrastructure. You’ll have to use relatively expensive 1G/10G ports to be able to connect the current and future servers, and once the server and virtualization engineers wake up and do hardware refresh you’ll end up with way too many ports (oh, and you do know that transceivers could cost more than the switching hardware, right?).

It was a dark and stormy Halloween night and a networking engineer was stuck in a data center facing a Mission Impossible project: replace a failing Cat6500 with a brand-new Nexus 7000. Shouldn’t have been a problem, if only the cables were labeled.

Every time I’m running a classroom version of my Designing the Cloud Infrastructure workshop, I start with a simple question: “Who has more than 2000 VMs or bare-metal servers in the data center?”

I might see three hands on a good day; 90-95% of the audience have smaller data centers… and some of them get disappointed when I tell them they don’t need more than two ToR switches in their data center.

Want to know what the difference between Virtual Chassis and Virtual Chassis Fabric is? How Local Link Bias works? How ISSU on QFX 5100 works even though the box doesn’t have two supervisor boards? You’ll find answers to all these questions in new videos describing Juniper data center switches.

People talking about long-distance workload mobility and cloudbursting often forget the physical reality documented in the fallacies of distributed computing. Today we’ll focus on bandwidth, in a follow-up blog post we’ll deal with its ugly cousin latency.

TL&DR summary: If you plan to spread application components across the network without understanding their network requirements, you’ll get the results you deserve.

Randall Greer left a comment on my Revisited: Layer-2 DCI over VXLAN post saying:

Could you please elaborate on how VXLAN is a better option than OTV? As far as I can see, OTV doesn't suffer from the traffic tromboning you get from VXLAN. Sure you have to stretch your VLANs, but you're protected from bridging failures going over your DCI. OTV is also able to have multiple edge devices per site, so there's no single failure domain. It's even integrated with LISP to mitigate any sub-optimal traffic flows.

Before going through the individual points, let’s focus on the big picture: the failure domains.