In the end, the central "clients" of any datacenter are the CPUs, Memory, & Disk Drives. James Hamilton of Amazon has written extensively on the topic - if we can narrow ourselves down to these central components, we will be getting somewhere. That is not to say that the efficiencies of infrastructure components isn't applicable (it most certainly is), but the true goal is to deliver as little energy as possible to the computing components and, in parallel, ensure the optimal environment for them to do their work (you'll see some metrics out there talking about useful work, etc)...the cooling, heat extraction, etc.
Cutting to the chase, the ultimate (IMO) shall reflect both the power source (carbon impacts), power delivered (transmission) to computing components (and all losses) and the power used to provide/maintain proper environment (this includes server fans ala JH's writings) - then we will truly understand how/where we can further optimize our Information Plants...
Richard Donaldson
CEO - Core4 Systems / 6connect
As Heather Clancy wrote yesterday in her “Thinking micro about PUE” piece, a vendor of energy management systems has proposed a much more low-level energy metric tied directly to power and cooling devices that it refers to as Micro PUE. The vendor describes Micro PUE as “the actual amount of energy used to cool 1 kW of IT load through a given cooling unit.”
There is a fair amount of interesting commentary in the white paper, but I’m of the opinion that the whole issue simply begs the question. PUE is becoming an accepted metric, because at the macro level, it does actually provide a broadly accepted measurement that, when the appropriate steps are taken to improve a datacenter’s energy efficiency, does provide a good starting point for the processes. The problem with starting at the micro level is that you run into “the forest for the trees” problem.
The concept of micro PUE seems to be more of a fine-tuning process. When you’ve taken a data center through all of the macro level changes that can be made to have a positive impact on PUE you then have the opportunity to take a look at the more micro changes that will have an impact. At that point you can more practically weigh the value of the potential changes at the micro level.
Starting at the micro level and working up rarely works well. And at its worst, the potential for individual PUE ratings on equipment will lead to all sorts of true but eventually valueless ratings. I have to admit my immediate thought was the analogy to performance car parts. Multiple-vendors making verifiable claims of power gains for their particular part, and customers expecting to be able to aggregate those performance gains by purchasing a bevy of performance parts, whose sum is far less than the total of its whole.
While I’ve oft discussed my concerns about the value of the basic PUE metric, I’m not of the opinion that a narrowly focused metric is the right way to go for the organization looking to make their datacenter more efficient.
