Environmental Leader

While I’ve been an advocate for this at points in the past, I’ve since come to the conclusion that the analogy is fundamentally flawed. The first issue is that there’s no good corollary to a ‘mile’ for server workloads – HPC, database, web, ERP all have different behaviors, so putting out one number doesn’t help if people can’t connect it to what they really want to do with the server. The bigger problem, though, is that the idea that there’s a ‘car’ is an illusion. The power in servers is used by processors, memory, disks and I/O cards, but your typical 2 or 4 socket server has multiple useful configuration options for each of these, resulting in a combinatorial explosion of possibilities where the lowest power and the highest may be 4x or more apart. Blade servers make this even worse. Finally, this leads to the last flaw, which is to imagine that someone purchasing racks of servers is analogous to someone browsing a car lot. Server purchasers, by definition, need to know what configuration they want and why. Anyone who’s taken that step is well equipped to deal with the real data, which is the impact that each of their configuration decisions has on the result, and to then compare the result to what they’d by from others. This is why most of the big suppliers have gone to detailed power calculators for their servers, so that customers can look at reasonably accurate, expected power for the actual configurations they’re interested in.

David Douglas | November 26th, 2008

The EPA is already working on this with their ENERGY STAR program. We did a webinar on this topic a few months ago: http://www.42u.com/energy-star-webinar.htm

In summary, the baseline data collection process is under way. Once they have the statistical base of data required they will use that to determine the ENERGY STAR rating, or MPG, for the equipment in the study.

Hang in there Jack, paraphrasing some well worn market speak; it’s coming soon to a data center near you.

Jacque Swartz | November 26th, 2008

Great article, Jack. Unfortunately it’s not that easy to agree on a single productivity metric for servers, because servers are designed differently to do different things. You and others are driving some terrific work on data center productivity in the Green Grid, but mapping that to individual server figures of merit — whether MIPS or CUPS or FLOPS or SPECxxx or TPCxxx — is difficult. Holding every server to the same productivity standard is like judging every vehicle only by its MPG rating. If we did this to all vehicles, fire trucks and ambulances would be legislated off the streets since they have really bad MPG. A typical data center needs servers of many kinds each designed for a different purpose. Trying to judge them all by a single metric leads to vendors trying to game the system, as the SPECpower initiative has shown.

Oh, and one more thing in response to the comment from 42U: The upcoming Energy Star for servers spec is not an MPG rating. It does not measure how many miles — or useful work — a server delivers per watt of energy. The Tier 1 Energy Star spec specifies power supply efficiencies, reporting requirements and server idle power — it does not specify a standard for work done by the server. Rather than measuring MPG, it specifies how much fuel your car can burn when idling at a red light. Later versions of the Energy Star for Servers spec will likely require some level of work efficiency, but the current version does not. See http://www.energystar.gov/ia/partners/prod_development/new_specs/downloads/servers/Draft3_Server_Spec_110408.pdf

Subodh Bapat | November 29th, 2008