Robert Cringely has made an impact with his latest piece from the PBS bully-tech pulpit. (catch it here http://www.pbs.org/cringely/pulpit/pulpit20051020.html and also the tech.memeorandum coverage here)
I design reliable power systems and data centers for a living.
In general terms, you don't need 3 to 4 times more electricity for cooling a data center than you need to power the electronic loads. The rough rule of thumb is 1 to 1. That means 1 watt of connected data center load will need 1 watt of electricity for connected mechanical loads. A watt of electricity does not automatically become a watt of heat; some of that electricity is converted to light and motion. Cringely overstates the actual (as opposed to the name-plate rated value) power consumption of the drive arrays by at least double. He also gets muddled up when he discusses data center floor space; mixing up total square footage with the square footage of the connected load space. He uses an incorrect approach to calculating "necessary" space to support "equipment" space. He makes some incorrect assumptions about the types of drives used in many of the arrays. The Cringely argument also completely overlooks load diversity in a data center. Even if 6,500 square feet required 330 Watts per square foot, the total equipment foot print over the remainder of the space would have a much lower density on average. We refer to stacked racks of blade servers and drive arrays as "spot" loads. You don't need to provision the entire facility at the load density of its densest loads; what you do have to do is provide the flexibility to power those loads in a range of locations and to provide appropriate cooling for installation conditions. It is true that load densities are starting to increase over the norms of the last few years, however.
It's worth noting that Cringely is treading down a path we've been on before and the evidence regarding actual data center electricity utilization demonstrates different usage patterns than what was predicted.
If you're interested in discussing data center electricity usage and design with me Robert, then call me at my Calgary office at (403) 541-6475 (Updated: 2006/12/10)
(c) copyright 2004,2005,2006,2007 by Bradford Gibson. Almost all rights reserved. Citation for journalistic, academic or weblog use is permitted with attribution.
Data Center Load Diversity
I have an internal client that is attempting to justify building a new data center because they are out of power - - out of power according to the enterprise management tool which is programmed with peak electrical demand for each device installed in the data center. The calculated demand exceeds actual demand by 200%.
I agree with running branch circuits to cabinets to handle peak loads, but deciding to jump ship before using 50% of installed redundant infrastructure seems a bit much. These data center folks consider the "zero risk" methodology to be the new model for managing data centers. Is it reasonable to presume that the new benchmark for managing data centers will be freeze when connected load equals UPS capacity and assume that all equipment will spin up and run at peak at the same time? How do I convince others to consider load diversity and where might I find resource material to help develop a concept model?
Post new comment