advertise
« Stuff The Internet Says On Scalability For April 22, 2011 | Main | 6 Ways Not to Scale that Will Make You Hip, Popular and Loved By VCs »
Wednesday
Apr202011

Packet Pushers: How to Build a Low Cost Data Center

The main thrust of the Packet Pushers Show 41 episode was to reveal and ruminate over the horrors of a successful attack on RSA, which puts the whole world security complex at risk. Near the end, at about 46 minutes in, there was an excellent section on how to go about building out a low cost datacenter.

Who cares? Well, someone emailed me this exact same question awhile back and I had a pretty useless response. So here's making up for that by summarizing the recommendations from the elite Packet Pushers cabal:

 

  • Look at Arista and Juniper. 
  • Juniper
    • Has a range of stackable switches, which includes some 10 gig.
    • If your budget can stretch for it they might make a good deal on their new QFX proto-fabric product. You can't get a full sized fabric solution, but you can get a few switches together to make a two port fabric. Good solution if you are running 10 gig and only need 30 or 40 10 gig ports. Thinks Juniper would make a good deal in order to get a few reference sites for QFX.
  • Arista Networks
    • Heard good things from financial institutions.
    • Also looking to get some business done.
  • Stick with multimode on your 10 gig, don't go singlemode.
  • Go for coax, but transceivers are expensive. 
  • Look for cabling solutions that are practical. Use click/clack cabling, don't use the pig tails that you fusion splice into the data center.
  • If power is an issue consider Arista as they make lower power equipment. 
  • Stackables vs chassis. Doesn't like stackables. Urination and burning was mentioned. For smaller datacenters a stackable can work, this assumes that because you have a lower cost you would also have a lower SLA and can absorb a few outages. The Juniper EX would be worth a look. The EX has 5 meter cables that can run between the switches for a 128 Gbps bus between the switches, which is good for the money. This is a nice design that scales out well for a low cost data center. 
  • The best way to manage maintenance costs is to buy it up front. When you make your first purchase for the datacenter tell your vendor that you want five years of maintenance. Cisco maintenance looks quite expensive at this point. For every $1000 of equipment you buy over five years you would expect to pay the same in maintenance. By buying it all upfront in one lump sum you can get significant maintenance discounts.
  • Lifetime support is really lifetime replacement support. You are still going to have to buy a maintenance contract. 
    • If you are going to rely on free support then the results will probably be equal to the cost.
    • Lifetime is the life of the model. It's not 20 years, it's as long as the model is supported.
    • When a model gets near its end-of-life you'll probably get a refurbished replacement from previously broken devices that had been returned and fixed.
    • Don't let limited lifetime warranties be your only drivers for cost. You'll assume more of the cost because you'll have to trouble shoot the problem to get a replacement. With support you could save that six hours of time letting support diagnose the problem.
  • Don't think about the vendor first. Get your requirements and your budget. Then go out into the marketplace and see who can meet them. Starting with vendors without requirements means you can be sold any sort of nonsense. Having requirements means you at least have what you need from the start.
  • Common mistakes: Get talked into road maps; Don't determine requirements. Best cost savings move you can make is to follow PML or SDLC life cycle. Determine requirements and map out where you plan to be in a few years. 
  • Keep-it-simple. You don't need those sexy new features. Focus on the requirements instead of device features to save costs. Focus on intangibles. What does your choice mean in terms of support? Training? What's the real ROI? The mistake made in this industry is we don't keep to the requirements.
  • A problem with using newer companies in what you gain in cost savings or with new features, you may lose on support. Maybe you buy a product that goes out of business and your product becomes an expensive door stop. You may not save what you think you are saving in the end.
  • Wouldn't buy Avaya because they are in Germany and have little presence in Europe. Known major bugs in equipment.
  • Wouldn't consider Cisco for a low cost buildout. Cisco likes to charge a lot upfront and a lot for maintenance.

Interesting stuff and a lot of useful tips. Hopefully my karmic balance has been restored.

Please listen to the podcast for the full effect and the disturbing implications of the RSA hack.

Related Articles

Reader Comments (4)

"Wouldn't buy Avaya because they are in Germany and have little presence in Europe." I know things happen so fast these days, but I am sure Germany is still part of Europe.

April 21, 2011 | Unregistered CommenterJames

Seems they completely skipped over the fact that the EX series doesn't require you to use their stacking cables. You can also build a VC together using fiber uplink ports configured for VC. They're absolutely on crack if they don't like "stackable" switches. I guess I can understand if they had Cisco in mind, though. If you're doing ToR deployment at a DC, the Juniper EX is by far the best option on the market - no question.

speaking of Cost effective datacenters, all aspects, not only networking. here is a post from outbrain tech blog.
http://techblog.outbrain.com/2011/04/lego-bricks-our-data-center-architechture/

April 21, 2011 | Unregistered CommenterOri Lahav

I'm surprised that Force-10 wasn't mentioned.

June 9, 2011 | Unregistered CommenterEllisGL

PostPost a New Comment

Enter your information below to add a new comment.
Author Email (optional):
Author URL (optional):
Post:
 
Some HTML allowed: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <code> <em> <i> <strike> <strong>