To continue the discussion of data center power consumption, I'd like to take a quick look at alternatives for addressing power utilization by optimizing the hardware side of the problem.

Before I begin, I wanted to share an interesting link  with you about the power consumption of each avatar (simulated person) in the online simulation Second Life.  I wanted to say it is amusing, but it actually raises worrisome questions about our whole IT infrastructure and who ends up paying these energy costs.  The article points out that the average avatar consumes about as much power as a Brazilian citizen.  You can read more here.  In any case, the article reminded me how important the link is from a company's business model through its software implementation, then through its hardware, then the electricity bill and finally the bottom line that investors will see.

As I mentioned in the first article, power supply efficiency has a large impact on server power consumption and heat generation.  If you are buying servers in small lots, it's hard to have control over what's inside your servers.  For example, the average server power supply is 65% efficient.  Compare this to Google, which custom-builds their own servers from selected off-the-shelf components and achieves claimed power supply efficiencies of 90%.  In ENKI's data center, we're also building our own servers to optimize the hardware environment for our business.  After optimizing for the cost of physical space for the server and equipment costs, we don't always have the flexibility to choose the highest efficiency power supplies.  However, whenever possible we choose 80+% efficiency power supplies.   For a 500-watt server, this translates to a 144-watt reduction, or about a $150 yearly savings (at an assumed electricity cost of about $0.12 per kilowatt hour.)  Recently a number of large vendors including Dell and IBM have started to deliver servers with power supply efficiencies of up to 91%, but at a price premium.  Given the inevitable electricity rate increases and the fact that you will also save on cooling costs, these servers can save you a significant amount over the life of the equipment.

 
If you're buying larger quantities of servers, it may make sense to buy them with DC power supplies which are much more efficient, often exceeding 90%.  A few vendors offer this option but before you go for it, make sure that you have DC power available and it is reasonably priced.  Many data centers charge a stiff premium for supplying DC power, including a provisioning fee.  Also, you'll need to rewire your entire power supply distribution system for the higher currents that low-voltage DC requires.  You'll want to balance these extra costs against the energy savings!  Another option is making your own DC power.  Rackable Systems, for example, offers AC-DC converters (rectifiers) that go in the top of a rack of servers, and supply redundant DC power to each server.   Rackable claims a total of 30% power savings using this configuration (implying that their overall power supply efficiency is close to 95%), as well as increased server reliability because a lot of the waste heat is confined to the rectifiers in the top of the rack.  Once again, you'll need to sharpen your pencil to determine if the considerable extra cost of the Rackable solution will save you money during your planning horizon.

 
After the power supply, the next largest source of heat in a server is the processor.  Both Intel and AMD claim to have lower power utilization than the other, and each has an interesting story to back it up.  The current Intel processors use less power than equivalent-performance AMD chips, but when you factor in the power from their "chipset" (external support circuitry and memory controller) and the special FB-DIMM memory they require, AMD's systems may use less power overall.  Sun's "Niagara" processor was specifically designed to save power, and some reviews show it to be three to five times more efficient than the latest AMD or Intel processors.  Between different rating metrics and measurement philosophies from each processor vendor, it can be hard to tell which will be more efficient; so many experts recommend that you actually measure the power usage of each server you are interested in buying under your typical loads to determine which has the lowest power consumption.  We use a "clamp-on" current meter and test our servers with actual application software loads.  In general, you can save power by increasing the number of processor cores in a server, since the power-hungry memory and support circuitry is shared.  Blade systems offer some of the highest processing power per watt efficiency, but because of the large number of processors per case, they also draw a lot of power so you will want to calculate the total power draw per rack before you assume your current infrastructure can handle the current.  Don't forget that with the large power density of blade solutions, power supply efficiency becomes even more important. 

 
Finally, there are things you can do to reduce incidental power consumption from cooling.  HP, APC, EMC, and IBM have thermal management systems and services that can make a new or remodeled data center more efficient by controlling the amount of cooling sent to each rack of equipment and hence reducing cooling requirements.  These are expensive options best applied to an entire large data center, but these vendors claim to be able to save you significant sums.  On a smaller scale, Rackable has optimized cooling in their servers and fully-configured racks containing back-to-back servers with passive airflow technology.  This saves power used to run cooling fans and reduce temperature variation in your data center by eliminating the ubiquitous hot aisle/cold aisle layout.  Chatsworth Products Inc. has some nice retrofit racks that can do the same thing with your choice of server hardware.

 
One thing to remember is that data center equipment has a finite life.  Depreciation rules allow you to set the financial life of a server, and now that manufacturers understand that power consumption matters, you can expect increases in computing efficiency to occur regularly over time.  Thus, it makes sense to rotate your equipment through your data center with regular purchases to keep the average power consumption down, much like a good investor would purchase stock monthly rather than all at once.  This means you will want to plan for any growth and make arrangements to only purchase what you need in each period.

 
In my next article on data center power consumption, we'll move on to software and systems architecture options to reduce power usage, including but not limited to virtualization.