The asymptotic race to the Power Utilization Efficiency (PUE) of 1.0

I think Mike Manos of DRT put it best – it is like an arms race to see who can achieve the lowest PUE. PUE is a metric used to evaluate today’s datacenter’s power efficiency by measuring a facilities total power consumption (X) and dividing it by the electricity of the “IT Load” (Y) or all the computing equipment. Some datacenters look at the power utilities meter for X and then either aggregate all UPS measurements (since they normally are not “networked”) or all the cabinet/rack PDU’s (both very manual processes). For you electrical engineers, this does not take into consideration any step downs from transmission rates or the multiple conversions to get to 240/120 inside the cab (nor the loss associated with server power supplies, etc).

The metric is typically expressed as some number that exists between 1.0 (theoretical nadir) and, the highest I’ve seen to date, 3.5. The EPA did a study and found, out of sample size of 100 datacenters, that the national annualized avg was 2.4 (I think this is very, very low) – this means for every Watt going to a server, 1.4 Watts is consumed by the facility (e.g., cooling, lights, monitoring, security, etc – anything not inside a computer rack). This metric is a far cry from the “Tier” system (Uptime recently bought by Group 451) as it starts to really hone in on what is important – isolating the “useful work” being done inside an “information plant” and trying to minimize all else…but it still has a way to go.

As evidenced today, we see people tossing PUE numbers around as if they are facts and not approximations (any PUE metric is a point in time – PUE is affected by too many things to remain fixed today). And the current thinking (folks like to look smart and toss around PUE as an indication of how much they know) is the lower the PUE, the better my datacenter…while true that a low PUE is a good thing (for above plus harmony with the environment and simply being more efficient), we are reaching a point where you can not go lower…unless you are generating power on site (co-gen) and then giving power back to grid (think hybrid cars), you will always have SOME power outside of your CPU/Mem/Storage that is required. Thus, what is that magic PUE nadir? 1.05? 1.1? 1.2?

This begins to sound a lot like the opposite race to achieve more and more reliability – 99.9% up; 99.99% up; 99.999% up; 99.9999% up – what we found was that somewhere around 4 or 5 nines was sufficient for our uptime as Murhpy proves to us, things will go wrong in this universe of entropy.

So while it is laudable to continually push the lower limits of PUE, we must not get caught up in it as the end all be all metric (for it is not) – it is a great evolution from Tiered systems (which you still see as part of RFPs today btw), but it still hasn’t been fully standardized upon for us to make apples to apples comparisons (one PUE is not like the other – all depends on assumptions at time of measurement). So as Mike Manos states, the datacenters’ “arms race” is to push us to “mutually assured efficiencies” which is far better for us all than the military version…:-)