Debunking the MPG myth...
If you saw the latest entry in the eco community blog (Eco Community Blog) you will get a good, but, not quite accurate, example of trying to define a standard metric. Here is the relevant snipet.
"A simple example: we don't concern ourselves with the efficiency of individual components of cars. Toyota doesn't market the Camry as having 80% efficient fuel injectors, or .12 coefficient air friction body styling (those numbers are made up). They just say 34 miles per gallon highway, 24mpg city (not made up). It's a measure of a given amount of work for a given consumption of energy, and flawed though those calculations may often be, they are at least comparable from one model to another, one manufacturer to another."
The basic idea is sound, however, as a self-proclaimed "car freak", it is not quite accurate and that has implications.
Let us look at the slightly flawed reasoning. MPG is not a gaurantee of the actual milage you will get. Its not a bad metric, but, an accurate representation in the real world, it ain't.
Ask yourself this question. Do you get the same MPG as your car is rated for?
Chances are probably not. It has a lot to do with your driving style, what you have in the car (weight), the lubricants used in the motor and transmission (friction), the air pressure in your tires (rolling resistance), the exact terrain and weather you are driving in (do you have a head wind or a tail wind) the outside air temp (the colder the intake charge, the better), and the octane of the gas you are using. (there is a reason why there are different octane levels, not to mention race gas.)
So your exact MPG will be different, depending on all these factors. So MPG is NOT measure of a given amount of work for a given consumption of energy. It IS a measure of a given amount of work for a given consumption of energy based on a number of variables, that they ain't telling you about.
Think of MPG as slightly less of a lie, or a slightly less half-truth.
So back to energy usage for a system in a data center. (And don't over look that word "system", it is the real crux of the matter.) It is the performance of the entire system (computer, storage, network, etc) that needs to be given. Saying "machine X uses 200 watts and machine Y uses 220 watts" doesn't mean anything unless the entire workload you are comparing on machines X & Y is totally contained on those single machines. No external storage, no network gear, no other systems operating as a different tier layer. If any of those do exist. Than the only accurate representation of the system MUST include all of those things.
Did you know that for every SPEC and HPC benchmark, a full disclosure of exactly what the HW and SW configuration was for that benchmark is required. (here are a couple chosen at random)
Specweb2005 results page
TPC-H results
Check the disclosure reports, there is a full spec on exactly what the specific system config was for each result.
So perhaps that is the answer. Can we get the benchmark standards groups to make power reporting (how much electricity is used on all the parts of a given config for a given benchmark) part of the reporting rules? Hook up a few power monitors to the configs and get the data and force the companies reporting results to publish that info as well.
I grant you, it wouldn't be perfect, but, it would be a pretty good starting point. It would tell a customer what the power usage was for a specific configuration, running a specific workload, what the result of that workload was, and how much power it took to do that work. Now that is a REAL metric, based on a fully disclosed set of parameters.
Posted at
10:13PM Mar 07, 2006
by rdsesq in Sun |