Last blog I talked generally about the relationship of Computation Output to Efficiency. In a world where energy growth is bounded, continued exponential growth in compute output will REQUIRE revolutionary changes in efficiency (the effectiveness of how energy is used).
The Industry has already made some huge progress in addressing efficiency of the data center. As James Hamilton recently wrote, the great thing about this industry is that when we focus, we get results. For instance The Green Grid (*) has been instrumental in driving among its partners on the use of PUE as a metric of data center efficiency has been astounding. While no metric is perfect, if used in the right way (as an incentive to decrease the a sizeable fraction of energy used to maintain the environmental conditions in a data center) it can drive substantial changes in data center energy consumption. Examples from Intel IT , Microsoft , Data Center 2020, and many other show how this has been put to use.
Another area where significant progress has been made is through the creation of the SPECpower_ssj2008 benchmark. The was really "a first" benchmark of its kind to look at the efficiency of computation in the way computer servers are actually used. Early on both "TDP" (often misunderstood - it is actually a thermal design power spec) and "Idle" (equally mistunderstood - what is an "idle" server?) were used as metrics of efficiency. While these are somewhat useful (there is no such thing as a perfect metric), SPECpower, which measures energy consumption at finite, but not full, workloads, does a much better job mimicking the energy a real user would see consumed.
A significant conributor to improving platform energy efficiency is the efficiency of power delivery. A few years ago, power delivery efficiency approaching 50% was common practice. Today, thanks to the effort of the Climate Savers Computing, some substantial progress has been acheived (for servers and for client systems). In 2010 alone, almost 6 Billion kWh will be saved because of this effort.
So we have a good start!
But the question is where can we go? I believe that ultimately optimizing the efficiency of data centers will involve optimization far beyond these two metrics. Below is a picture I drew showing a "STACK OF ENERGY USES" in the data center. Each block represents the "next level of loss" in the chain of energy consumption all they way down to the gates of the tranististors doing the calculations (the work output of the data center).
This is a crude picture and I am sure people will have comments. I've highlighted the energy usages that the two metrics I mention above seek to optimize. As I stare at this, I am wondering what other kinds of metrics should be put into place to improve energy usage, and why? Do the duo of SPECpower and PUE provide ample indicators to drive the right behaviors?
In my next blog I'll weigh in with some ideas.
*as a disclaimer, I am an active participant in The Green Grid and an alternate on that organizations Board of Directors.