First there was SETI@home where you could use spare CPU cycles to look for extraterrestrial signals. Now we have Climate@Home, running NASA GISS modeling software in distributed computing, but no release date yet. Oxford has already created climateprediction.net running BOINC. I expect it’s a case of “keeping up with the Joneses”
I wonder though, what’s the “carbon footprint” of leaving all those computers running to calculate the effects of fossil fuels burned to run them? Already the Met Office Supercomputer is drawing fire for it’s massive electricity use, so does this simply spread the problem around?
From NASA’s CIO: Creating a Virtual Supercomputer to Model Climate
NASA will be calling on people worldwide to help determine the accuracy of a computer model that scientists use to predict climate change. The initiative, called “Climate@Home,” is unprecedented in scope. Never before has NASA attempted to recruit so many people to help perform research vital to forecasting the Earth’s climate in the 21st century under a wide range of different situations.
NASA’s Earth Science Division (ESD) and Office of the Chief Information Officer (OCIO) have strategically partnered to manage the Climate@Home initiative. This effort will include collaborations between the 10 NASA Centers, the 13 Federal agencies of the USGCRP (United States Global Change Research Program) along with several universities and private organizations.
Goddard Space Flight Center (GSFC)’s Robert Cahalan is serving as the project scientist and has assembled an international team of scientists to help set science goals and determine which parameters to run. GSFC’s senior advisor to the CIO, Myra Bambacus, serves as the project manager and will run this initiative.
Participants need no special training to get involved in Climate@Home. All they need is a desktop computer or laptop. Volunteers will be able to download the computer model to run on their computers as a background process whenever the computers are on, but not used to their full capacity.
The climate model that volunteers download is made up of mathematical equations that quantitatively describe how atmospheric temperature, air pressure, winds, water vapor, clouds, precipitation and other factors all respond to the Sun’s heating of the Earth’s surface and atmosphere. Models help predict how the Earth’s climate might respond to small changes in Earth’s ability to absorb sunlight or radiate energy into space.
Scientists traditionally have used supercomputers to test the sensitivity and accuracy of climate models. With these powerful machines, they are able to run millions of calculations, each computing a different scenario or combination of circumstances such as varying levels of chlorine or water vapor in the atmosphere.
Instead of using supercomputers in this effort, NASA is creating a virtual supercomputing network that spreads the data-processing chores across thousands of computers. Such a task-sharing initiative eliminates the need to buy additional supercomputers, which consume enormous amounts of energy, and reduces the “carbon footprint” of running these calculations.
Prior to the initiative’s official roll out in early 2011, the project will be announced in the news. The project website will provide instructions on how to download the models and supporting computer software. The goal is to have recruited tens of thousands of participants by the time the initiative begins. Each participant will run the same model but with certain parameters slightly adjusted. Scientists will examine the resulting sensitivity of the climate predictions to those adjustments, resulting in a better understanding of the parameters that should be studied in the future.
Climate@Home will have the option of using the same high-level architecture developed for “SETI@Home,” a scientific experiment that uses Internet-connected computers in the Search for Extraterrestrial Intelligence program. The initiative also is modeled after a similar project coordinated by the Oxford e-Research Centre in the United Kingdom called Climateprediction.net.
Climate@Home will test the accuracy of a model developed by the Goddard Institute of Space Studies (GISS) in New York, and will serve as a trailblazer to explore the accuracy of other models as well.
==============
h/t to WUWT reader Bruce Foutch
I have chosen to help out the Milkyway@Home project instead. It is mapping out a “highly accurate” 3D model of the galaxy. More info at http://milkyway.cs.rpi.edu/milkyway/ — John M Reynolds
” Paul Linsay says:
August 25, 2010 at 6:25 pm
I did the ClimatePrediction.net thing. It got withdrawn because about half the time the model would plunge into an Ice Age. Can’t have that, can we?”
That sounds like the closest thing we have to an accurate model.
There is a business out there for programmers!: Make a Climate w/opt. Warming/Cooling Model for phones
The main thing for people to understand about this project is that the climate “model” in question is NOT running as a parallel “supercomputer”. That is, this project has nothing to do with running one large (high resolution) “model” in parallel, where small portions of the computation are divided among thousands of processors (as is the case with a true parallel code running on LINUX and Windows clusters). Rather, this is a case of taking a small, cruddy little computer “model” and using someone’s PC to run one or more cases. Software like BOINC basically coordinate sending you the prepackaged executable and input data, your PC runs the code, and the output data is mailed back to the coordinating entity.
I don’t know what they are thinking in terms of proving “accuracy”. Again, accuracy relative to what? And there’s no way you’re going to show any kind of grid convergence (PCs with a couple of gig of RAM are way too small) of the “solutions” either, not that they worry about such things at GISS.
This is, unfortunately, yet another boondoggle to waste more taxpayer money…
BOINC’s climateprediction.net is what got me interested in looking into the science behind AGW. After reviewing the scientific evidence, I quickly disposed of climateprediction.net, and donated my computer time to Rosetta and World Community Grid. Both of these projects study the molecular structures of diseases (either the disease itself, or how drugs interact with various viruses or bacterium).
Ironically, it is climateprediction.net that led to my eventual discovery of WUWT!
However, don’t throw the baby out with the bath water on this one. While SETI@home, climateprediction.net and climate@home may indeed by “junk science”, there exists a wealth of worthier causes that you can painlessly help with on the BOINC network. Check out: http://boinc.berkeley.edu/projects.php for a short list of projects.
Q – How do you enlist hundreds of millions to your CAUSE and PERSPECTIVE and RELIGION?
A – Make everyone of them feel as though they were part of the discovery of something NEW, REVOLUTIONARY, CRITICAL. (Glooooooooooooobal Warrrrrrrrrrrrrrrrming!!!!!!!!!!!!!)
I never took to fancy SETI@Home, partly because I didn’t want no super weirdo nerds crawling around my computer, but mostly because I couldn’t see the whole intelligent aspect of looking for highly intelligent life, i.e. other then monkeys, dolphins, and gold fish, nor the über cleverness of connecting potentially millions of computers digitally communicating with each other to find that highly intelligent life by trying to find patterns in radio waves of analog nature.
But sure you guys go head and waste more money on an otherwise already too high electricity bill. :p
No way I’m doing that! It will make my computer run too hot and likely burn out the fan.
Sounds like a great way to find the set of parameters that cooks up the scariest scenario to report to the press.
pwl: August 26, 2010 at 12:12 am
“…I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.”
– Charles Babbage, The Inventor of the Analytical Engine – the first computer
Because people believe that computers make conscious decisions. They have never been forcefully reminded that a computer is only a machine that can add and subtract 1s and 0s faster than they can.
Ah people..
testing a model for sensitivity to input parameters EVEN BOGUS input parameters is STANDARD practice, REQUIRED practice. Whether one is building a model that predicts how a plane will fly ( they are “wrong” but we use them) or how a car will drive.
And, we often use a hierarchy of models from the simple to the complex, but they all should be tested for sensitivity to input parameters. Like SO:
What happens if I assume that we got levels of methane wrong? I dunno, run the
model with 100 different levels.. how sensitive is the model to THAT input.
not very sensistive? or very sensitive? its the only way you can understand how
big models perform.
What they are doing is actually good practice
The BBC beat them to it (carried out in 2006)
http://www.bbc.co.uk/sn/climateexperiment/
I suspect that the results will be similar (Predicted impact on the UK)
http://www.bbc.co.uk/sn/climateexperiment/whattheymean/theuk.shtml
Steven Mosher says:
August 26, 2010 at 9:57 am
“What happens if I assume that we got levels of methane wrong? I dunno, run the
model with 100 different levels.. how sensitive is the model to THAT input.
not very sensistive? or very sensitive? its the only way you can understand how
big models perform.”
“What they are doing is actually good practice.”
Well, OK, if it is good practice:
(1) Why haven’t they done it already? The model they probably will use is of 80 – early 90s vintage. If it will run on a PC, surely they have the computing HP to do this themselves, well before even the first IPCC report. By they way, didn’t GSFC give these people a new $5 million upgraded computing cluster for this kind of work (using stimulus funding no less)? Why knot use that? (That is if they can get GISS personnel to stop blogging long enough to work on it…).
(2) Knowing that simple-minded, low-res model A has a certain output behavior to a given input does NOT imply that your “big” model B will behave accordingly. Unless, of course, it’s based on the same algorithms, numerical formulations, and coding practices (unlikely – and with GISS unknowable). So, you can’t use the results of this exercise to claim the recently numerical solutions from some version of Model E will act similarly – they are two different codes.
And the cost of this boondoggle is NOT zero. The compute cycles may be “free” but there will need to be a lot of work (=money) to organize, check, and publish the findings. We simply DO NOT HAVE THE MONEY ANYMORE…
This is a brilliant strategy by NASA.
When the climate prediction numbers start coming in wrong with this distributed computing, NASA can say: “It wasn’t us. It must be a defective laptop out there.”
I can help. I already have a therometer in the shade, and the readout is in the bathroom, just under the attic door. The attic temp should be fairly close to the temp they are looking at when they look at my roof and tell everyone we’er in global warming.
Too bad you can’t upload a “reality” virus to GISS via BOINC.
There is only one system in the whole of the known Universe capable of modelling the Earth’s climate.
We live on it.
Anybody that thinks that an open-ended non-linear system with a practically infinite number of feedbacks, most of which we don’t even know yet, and even the ones we do we aren’t sure of the sign of (take clouds for example, where the sign alters continuously according to whether it’s day or night) is capable of telling us anything whatsoever about the future climate either doesn’t know anything about science – especially computer science – and mathematics, or is a plain downright liar.
Or a fool, of course.
Garbage IN,
IPCC Fifth Assessment Report-acceptable peer-reviewed literature OUT.
Wait for it, it’ll happen.
@Catweazle
‘Anybody’ […] ‘capable of telling us anything whatsoever about the future climate’ […] ‘is a plain downright liar.’
‘Or a fool, of course.’
Sorry I had to edit your statement, but I must say for being a fool I’m quite happy at telling you anything what so ever you want to here, especially for a price! :p
Frank:
“(1) Why haven’t they done it already? The model they probably will use is of 80 – early 90s vintage. If it will run on a PC, surely they have the computing HP to do this themselves, well before even the first IPCC report. By they way, didn’t GSFC give these people a new $5 million upgraded computing cluster for this kind of work (using stimulus funding no less)? Why knot use that? (That is if they can get GISS personnel to stop blogging long enough to work on it…).”
1. Why would you assume that they havent done it.
2. why would you assume that sensitivity testing is EVER complete.
3. The cycle hours available in a year of computing are known. They know
how many runs they have to do for Ar5. You prioritize your use of the assets.
So, one can well imagine that they want to do less important sensitivity studies
using off campus cycles.
“(2) Knowing that simple-minded, low-res model A has a certain output behavior to a given input does NOT imply that your “big” model B will behave accordingly. Unless, of course, it’s based on the same algorithms, numerical formulations, and coding practices (unlikely – and with GISS unknowable). So, you can’t use the results of this exercise to claim the recently numerical solutions from some version of Model E will act similarly – they are two different codes.”
Well, you dont understand how hierarchies of models work. There are many ways to benchmark low res models to high res. I’ll give you one. In determining what level of RCS ( radar cross section) a future fighter requires, very often we would use low res models to BOUND the the problem or bound the sensitivity study space. So for example, we would study very small RCS and then vary that parameter upwards looking at the response curves ( usually for non linear behavior) then using a few point off that curve we would run the high res models and benchmark.
So for a low res model we might varying the parameter from 0 to 100000. And you might find out that the parameter was only sensitive within 0-50. At values beyond 50 you got the same answer.
Well you take that information and first thing you do is check that assumption. The threshhold is 50. So with high res you might check
0,10,20,30,40,50,60, 70,80,90,100,200,300,400,500,1000,2000, 20,000, 50,000..
you might have computer time for 100 runs.. so you use low res to pick your window.
you are not concerned with exact matches in the response value
So , now, hopefully you get a confirmation of the systematic behavior. maybe the high res has a threshhold at 100 as opposed to 50.. ANd you note that changes from 0 to 25 do nothing!
basically you use the low res models to get an idea about the response characteristics of various input parameters. That reduces the paramter space you have to look at. So next time, you have 100 runs to split up in the parameter space.
so you might do.. 0, 20,21,22,,23,24, 25,26,27,28,29,30,35,45,50,55, ETC. and you expand the number of high fidelity runs you do around the past break points
Steven Mosher says:
August 26, 2010 at 6:01 pm
Steve – I’ve been studying and implementing computational fluid dynamics codes for well over 20 years. I understand quite well how hierarchies of models work.
With all due respect, your example, I believe, is not valid in this case as we’re talking about systems of non-linear, coupled, partial differential equations. The formulations, algorithms, numerical implementation, parameterizations, and coding are all very crucial to the final result. The fact of the matter is we’re talking about using PCs to examine the sensitivity of a small, crude, non-linear climate model, the internal workings of which bear little to resemblance to the large, complex models in use today.
Here’s an example for you. Let’s say we look at external flow over an obstacle at a Reynolds number of 40,000. A crude model will be quite stable (dare I say robust?) and in fact yield a steady solution. Of course, as you start resolving the wake (with a finer mesh) you find that the solution no longer converges to steady-state, and in fact is highly unsteady. As a result, important quantities such as mean lift and drag coefficients can be way off the mark in the crude model.
And the above example is just a “simple” fluid dynamics problem! What if we now add cloud thermodynamics, tracers, radiation models, coupled ocean models, ice models, etc., and then attempt to integrate this complex, non-linear system over years and years of computational time? No one even knows if this problem is well-posed mathematically.
This current effort by NASA is just a colossal waste of time and money – money well simply do NOT have. The results will only be valid for the crude model on which the tests were conducted, and will be even more worthless given that I doubt we will even get know what differential equations are being solved! Unless, of course, NASA decides to finally properly document their codes…I’m not holding my breath…
SETI and CAGW have a lot in common – they’re both junk science.
That is, they formulate their theories in terms of parameters which can’t be physical measured (the theories can’t be falsified in the sense of Popper.)
In fact, it can be argued that SETI sanitised junk science for the public consumption.
http://sms.cam.ac.uk/media/871991;jsessionid=092E2588B5A7AF04E0F7B86007423826?format=flv&quality=high&fetch_type=stream
More ‘computes’ will never repair bad data and lousy programs.
The idea that more computes will improve their accuracy is fundamentally flawed. It may improve their precision, but not the accuracy. Better data and better code is needed for that.
In fact, the improvement in total “computes” from improved algorithms is more than that from improved hardware. The necessary corollary of that is that BAD algorithms can consume all available improvement in hardware, and then some, while producing nothing of value.
More hardware is only justified AFTER all the software and techniques questions are fixed.
With that said: I do LOVE Big Iron. I’ve managed a supercomputer center before and it is ‘way cool’… One of my favorite machines was the “Stonesouper” computer, a Beowulf Cluster made from cast off PCs. It inspired me to make my own little wulf from a half dozen boxes; one of which became the box I ran GIStemp on for the first time. I’d expected to need a lot of power, so did the port to one node of the cluster so I could add however nodes it might need. I was very surprised to find that ONE was more than enough. That box is now ‘semi-retired’ and I’ve got GIStemp running on a much newer HP Vectra now.
http://www.extremelinux.info/stonesoup/
http://chiefio.wordpress.com/2010/08/09/obsolescent-but-not-obsolete/
FWIW there are a variety of problem types characterized by how much the computation can be compartmentalized and how much communications must happen between the parts. The very highly distributed computing on platforms connected by very slow networks only really works well on highly modular problems. This implies their ‘model’ has a lot of disjoint cells that do not interact much with each other. Doesn’t sound like a really great approach to a climate model to me.
For an 8,000 cell model of the world it would be better to have an 8,000 “core” machine with fast interconnects than to have 8000 widely scattered desktops. Such machines are fairly common. (A 2000 CPU machine made with ‘quad core’ processors would do it) and would let the models run much much faster AND with much more interaction between the “cells”. I’d expect that with the interactions between the elements of weather one would need such interactions to have a valid model. If you wanted 16,000 cells, you could go to more processors, or just more memory and run the two groups in alternate time slots.
One Large CPU does not work as well for problems with lots of small cells. It works better for large arrays of similar data. One could probably code the models to work with large arrays and run a giant mondo processor on it, but that would not be my first choice of approach.
Bottom line, though, is that if they are going for the highly distributed, time asynchronous, slow communications channel approach to modeling, there will either be fairly small interaction between the cells (certainly small iterative interactions) OR they are using a very sub-optimal choice of hardware.
You could get out of that box by handing over a grid of, say, 64 cells and computing all their internal interactions and feedbacks, then “stitching it together” at the margins with the ‘neighbor’ 64 cell blocks and calculating their interactions; but that would be a bit of a pain (and less interesting than having a simultaneous solution for all the cells.) Basically: Yeah, you could make it go; but it would not be the best way to do it.
SETI works because each work unit is entirely independent of the neighbor work unit, so all computes can be done independently. A space alien in Alpha Centauri has little to do with one in Betelgeuse, so low computer inter-cell communications is fine (as there is zero inter cell dependency) But a change of ocean temps in Hawaii DOES have an impact on cells in Siberia, so the cells ought to have a fair amount of interprocess communications.
Somehow I smell a PR stunt more than a technical solution optimization.
Then again, it can be a very cost effective solution even if it is a poor one technically.
Frank K. says: No one even knows if this problem is well-posed mathematically.
It isn’t.
http://chiefio.wordpress.com/2010/07/17/derivative-of-integral-chaos-is-agw/
The models run off input data that are insufficient (Nyquist and quality both), turned into results that are meaningless via the temperature creation series such as GIStemp, then the models turn that into fantasies based on non-convergent math by playing forward a time series that diverges over time.
I’ve tried to get more folks interested in the math of the issue (not the arithmetic) but most folks just don’t care about the math issues. But they make the whole process just an exercise in dancing in the error bands of computer fantasies.
Oh, and per the idea of all the sensitivity testing being useful: It is IFF you know what the correct value is supposed to be. If you presume CO2 ought to be a strong positive feedback when it isn’t, doing sensitivity testing to find that yes, you got CO2 set it High Gain is not assessing the accuracy of the model, just the precision with which you have got the assumptions wrong.
FIRST get the models in touch with reality, THEN worry about how sensitive they are.