loopback0 – Douglas Gourlay's Blog Data Centers, Virtualization, and Cloud Computing

Validating Some Power/Cooling Cost Assertions

What is the easiest way to account for space/power consumption of a network element?

Am making a spreadsheet comparing different products and looking at longer term costs, maintenance, power, cooling, etc.  I felt that rather than scrubbing the DOE sites and trying to get power costs by state I would just use the national average, but then fell flat on that because I found negotiated rates could be much less than published tariff rates.

Then I stumbled upon what may be an easier solution to my quandary and one inline with what I see a lot of enterprises doing - call a hosting company.  I haven't talked to too many enterprise customers that are not at least considering if not seriously considering using a hosting environment, or event a full-blown cloud deployment for some portion of their enterprise data center workloads.  Why? - the main reason I keep hearing is that most enterprise customers cannot build big enough to achieve the same economy of scale as a Google, Microsoft, Facebook, etc.  So they may as well lease space from a provider who can achieve a higher density, lower PUE, better delta-T, and handle the compliance tasks like SAS 70 Type-II (Switch, Equinix, Corelink, etc) and not to mention the IT assets put within the data center grow at a power/performance curve that usually breaks the facility they are housed within in 5-7 years, so who wants that on their books - better to let the provider manage/operate it.

In asking around I got to an average number of ~$155 per month per kilowatt consumed when in a hosted environment (ping, power, pipe).  Does this seem inline to you or too high/low based on what you are seeing?

With this data you can then extrapolate Watts/10Gb port across several systems and you get variability from $92/year per 10GbE port up to $372/year per 10GbE port assuming $155/month per kilowatt.  (I am eliminating my own companies products from this so as to avoid being a blatant advertorial...) Annualized hosting/power cost comes to $9,400 to $25,800.

I will be the first to admit there are HUGE differences in features, programmability, buffering, network segmentation, encapsulation methods, and Quality of Service Granularity between many of these platforms.  Those that performed the best were usally more 'switch like' with smaller buffers, less features, and fixed function ASICs for the data path.  Those at the top end of the spectrum were almost always products like Juniper's T640/T1600 and Cisco's CRS - extremely high function core routers with huge performance, buffers, shapers, policers, and probably most importantly a software upgradeable packet processing engine that allows incremental feature additions that execute in the data plane.

It's clearly not an apples-to-apples and don't want it to come across that way, my real question is - is using an average of US hosting pricing per kilowatt an effective way to get a model for opex cost/10Gb port or are there other models people would recommend?  Am pretty open to anything right now provided it is accurate and neutrally intentioned.

dg