Four-Socket Power7 Boxes Get Energy Star Rating
April 12, 2010 Timothy Prickett Morgan
Last May, as The Four Hundred previously reported, the U.S. Environmental Protection Agency and the Department of Energy, along with two and a half years of input from the IT industry, established an Energy Star power efficiency benchmark and rating system for servers. That initial spec was for what was called computer servers, meaning single-socket and two-socket boxes. In February of this year, the spec was expanded to include what are called enterprise servers, which means machines with four or more sockets.
As it turns out, the new Power7-based Power 750 machines and their special supercomputer variant, the Power 755s, are the first four-socket boxes to be put through the Energy Star certification process. You can see the results of IBM’s Energy Star certification tests at this link.
As I explained last May. The rigorous monitoring of power usage by systems of similar configuration is extremely useful, but the Energy Star spec does not require a specific benchmark test to gauge performance consistently, so performance per watt comparisons are only possible on machines running the same test to load up the system. IBM used the Linpack Fortran supercomputer test on the Power 750 and 755, which is not exactly relevant to general-purpose computing that we see in data centers. On some two-socket System x machines IBM tested the machines using the Steam memory bandwidth test and something called the Multi-Platform Exerciser (or MPx) utility version 1.266.
The interesting thing that the tests on the Power 750 and 755 show is that as you load up the machine with more memory and disk drives to scale performance, the performance per watt drops rather dramatically. The power draw from extra disks and memory is higher than you might expect, roughly doubling power draw when going from a minimum to a maximum configuration, but only goosing performance a tiny bit. The lesson that one can draw, obviously from a very skinny bit of data, is that you need to carefully size your systems if you want to be as efficient with the juice as possible. Loading up systems, despite all the talk about virtualization and server consolidation, might not be as efficient as having clusters of more energy and correctly configured boxes. Yes, this runs contrary to logic. The point is, a midrange configuration is not much worse in terms of power draw than a minimum system, but a maxxed out system is nowhere near as efficient as either.