Category: data center

Andrew wondered how one could scale the L3-only data center networking approach I outlined in this blog post and asked:

When dealing with guests on each host, if each host injects a /32 for each guest, by the time the routes are on the spine, you're potentially well past the 128k route limit. Can you elaborate on how this can scale beyond 128k routes?

Short answer: it won’t.

Tore Anderson has been talking about IPv6-only data centers (and running a production one) for years. We know Facebook decided to go down that same path… but how hard would it be to start from scratch?

Not too hard if you want to do it, know what you're doing, and are willing to do more than buy boxes from established vendors. Donatas Abraitis documented one such approach, and he's not working for a startup but a 12-year-old company. So, don't claim it's impossible ;)

One of my readers sent me a lengthy email asking my opinion about his ideas for new data center design (yep, I pointed out there’s a service for that while replying to his email ;). He started with:

I have to design a DR solution for a large enterprise. They have two data centers connected via Fabric Path.

There’s a red flag right there…

Let’s continue our journey toward two-switch data center. What can we do after virtualizing the workload, getting rid of legacy technologies, and reducing the number of server uplinks to two?

How about replacing dedicated storage boxes with distributed file system?

In late September, Howard Marks will talk about software-defined storage in my Building Next Generation Data Center course. The course is sold out, but if you register for the spring 2017 session, you’ll get access to recording of Howard’s talk.

Remember our journey toward two-switch data center? So far we:

• Virtualized all workload we could;
• Removed legacy technologies.

Time for the next step: read a recent design guide from your favorite hypervisor vendor and reduce the number of server uplinks to two.

Not good enough? Building a bigger data center? There’s exactly one seat left in the Building Next Generation Data Center online course.

Three years ago I was speaking with one of the attendees of my overlay virtual networking workshop @ Interop Las Vegas and he asked me how soon I thought the overlay virtual networking technologies would be accepted in the enterprise networks.

My response: “you might be surprised at the speed of the uptake.” Turns out, I was wrong (again). Today I’m surprised at the lack of that speed.

One of my readers working as an enterprise data center architect sent me this question:

I've just finished a one-week POC with Arista. For fabric provisioning and automation, we were introduced to CloudVision. My impression is that there are still a lot of manual processes when using CloudVision.

Arista initially focused on DIY people and those people loved the tools Arista EOS gave them: Linux on the box, programmability, APIs… However

In our journey toward two-switch data center we covered:

• The overall idea;
• The need for high-density server virtualization (most of the times).

It’s time for the next step: get rid of legacy technologies like six 1GE interfaces per server or two FC interface cards in every server.

Need more details? Watch the Designing Private Cloud Infrastructure webinar. How about an interactive discussion? Register for the Building Next-Generation Data Center course.

Occasionally I get feedback that makes me say “it’s worth doing the webinars ;)”. Here’s one I got after the layer-2 session of Leaf-and-Spine Fabric Designs webinar:

I work at a higher level of the stack, so it was a real eye opener especially with so much opinionated "myths" on the web that haven't been critically challenged such as [the usefulness of] STP.

There’s more feedback on this web page where you can also buy the webinar recording (or register for the next session of the webinar once they are scheduled).

One of my readers left a comment on my “optimize your data center by virtualizing the servers” blog post saying (approximately):

Seems like LinkedIn did it without virtualization :) Can enterprises achieve this to some extent?

Assuming you want to replace physical servers with one or two CPU cores and 4GB of memory with modern servers having dozens of cores and hundreds of GB of memory the short answer is: not for a long time.

A Technology Market Builder (in his own words) from a major networking vendor decided to publish a thought leadership article (in my sarcastic words) describing how Cisco’s embrace of complexity harmed the whole networking industry.

Let’s see how black this kettle-blaming pot really is ;), and make sure to have fun reading the comments to the original article.

An engineer working for a large system integrator sent me this question:

Since you are running a detailed series on leaf-and-spine fabrics these days, could you please suggest if following design scenarios of Facebook and Linkedin Data centers are also covered?

Short answer: No.

One of my readers has customers that already experienced performance challenges with Tomahawk-based data center switches. He sent me an email along these lines:

My customers are concerned about buffer performance for packets that are 200 bytes and under. MORE IMPORTANTLY, a customer informed me that there were performance issues when running 4x25GE connections when one group of ports speaks to another group.

Reading the report Mellanox published not so long ago it seems there really is something fishy going on with Tomahawk.

Here’s an interesting story I got from one of my friends:

• A large organization used a disaster recovery strategy based on stretched IP subnets and restarting workloads with unchanged IP addresses in a secondary data center;
• Once they experienced a WAN connectivity failure in the primary data center and their disaster recovery plan kicked in.

However, while they were busy restarting the workloads in the secondary data center, and managed to get most of them up and running, the DCI link unexpectedly came back to life.

A month ago I published the video where I described the idea that “two switches is all you need in a medium-sized data center”. Now let’s dig into the details: the first step you have to take to optimize your data center infrastructure is to virtualize all servers.

For even more details, watch the Designing Private Cloud Infrastructure webinar, or register for the Building Next-Generation Data Center course.