Slashdot Mirror
Earth Simulator Now Predicting Hurricanes?
GeoGreg writes "The BBC is reporting that the Japanese Earth Simulator supercomputer is producing results showing that it is possible to model climate down to the level of severe weather events such as hurricanes. This computer has been discussed on Slashdot previously, and it sounds like at least some of the hype around this beastie was justified."
Model the destruction to the USA east coast that WILL happen when a large chunk of the Canaries falls into the sea, ( estimated to be sometime in the next few centuries ) now if the EU really wanted to 'influence' US policy a few studies like these, plus a bit of mining might do it ..... conspiracy theorists take note!
As mentioned before, there is a distributed project called climateprediction.net
for those who want to participate themselves. It is run by the University of Oxford in the UK, it is not affiliated with . So far only a windows client, but a Linux one is in the works. It is very CPU intensive, so if you have less than an 800mhz processor you shouldn't bother, it would take months to finish a single unit of work.
Hey, do you find it suprising that the nation that knows the least about climate science is the one that is most skeptical about global warming?
"In little more than a decade, the United States has fallen significantly behind other countries in its ability to simulate and predict long-term shifts in climate, according to a wide range of scientists and recent federal studies."
"During the Clinton administration, the lack of American modeling leadership did not have a discernible impact on climate policy, various experts said. But it did prevent the United States from playing a more central role in writing critical sections of the Intergovernmental Panel's report -- particularly the part assessing the extent of human influence on the warming trend of recent decades.
In computing power, Dr. Sarachik said, "our top two centers together don't amount to one-fifth of the European effort."
In that article from the New York Times is from two years ago! It mentions the japanese plans to build the Earth Simulator.
Being bitter is drinking poison and hoping someone else will die
I think it should be noted that the article says they can only predict the likelyhood of hurricains occuring, rather than actual individual instances of hurricains as that title of this story implys..... We wont be able to fast forward the model, and predict a hurricaine years before it comes along for instance.... (Somthing to do with Chaos theory ends all hope of this ever happening..... You'd have to predict a leaf falling someplace else first!).
While it is nice to know the computing power is out there, people need to realize that the prediction is only as good as the software the scientist. There are a lot of things that go into the weather. I question if we have caputred data on enough of them to really start making such long term predictions. I am curious if they have actually been able to modle past weather based upon the data they would have had avalable. Predicting the weather for what happened a year ago would be a neat trick, but only if you don't cheat and use more data then what would have been avalable if one had done it for real.
I wonder if the system releases only one pattern that the weather will follow or if it returns many different ways that the weather could go.
I would hope so. The National Center for Environmental Predictions (NCEP) does this now with their model called the Global Forecast System (GFS) that goes out to 384 hours or 16 days. With this model, they do something called ensemble forecasts where they rerun the model another ten times at a reduced resolution from the master run with perturbations added to each. Then they compare the results and will, on some of the graphics, use all ten to perform a type of averaging to remove the really bogus forecasts.
My experience has been if you are doing any type of long range forecasting, the ensemble method is the way to go. I am not saying that it is exact, but has proven an invaluable guide past day 4 for good long range forecasting. My guess is that this project in Japan would be taking this into account and performing something like this type of ensemble method. If not, I would seriously question their results.
Bryan R.
The price of freedom is eternal vigilance, or $12.50 as seen on eBay.....
Wow, I'm surprised that a resolution of 10 cubic kilometers is enough to actually make any predictions besides the most general of weather trends.
Define "the most general of weather trends". Currently, at least for here in the US, the model of choice (something always of debate) is the Eta that typically is run at 44km (they have much higher resolutions, but those aren't as readily available). Believe it or not, this model has been great at forecasting for frontal based weather (like thunderstorms along a cold front) and winter storm systems (it is able to place the areas of heavy snow by county) Depending on how close the model run is the the event, the placement of this information is usually pretty close.
That isn't to say it is perfect. As you could imagine for a grid that size, the model will typically miss popcorn type showers and thunderstorms. Also, if you do any severe weather forecasting, you will miss the small scale features like a tornado or such.
They have something called the RUC which is run at 20km. I am not as familiar with this model, but a person I work with has used it to do tornado forecasting (check out the historic data towards the bottom) and has had incredible results.
Bryan R.
The price of freedom is eternal vigilance, or $12.50 as seen on eBay.....
Sometimes you feel like a nut, and sometimes you don't... Sad really.
Keep in mind the top500 list is built by volountary submissions to the archive. That means - there is a not insubstantial number of computers that "should" be in the top 500 (and top 5 for that matter) but aren't because of either a) the researchers who bought the system are just itching to use their very expensive machine for research and just bypass running the benchmark to get right down to business b) the agency/company using the machine desires to maintain a low profile.
Also - keep in mind that linpack favors machines with very low latency, high bandwidth interconnects. But - in some simulations / applications this is irrelevant. Some applications demand this bandwidth/latency - but others transfer relatively small, but very difficult problems among CPU's. (SETI @ HOME is a decent laymen example of this... A work unit is quite small, but requires a lot of cpu time to process.) Thus actual performance of an application like this would more closely approximate the Rpeak figure rather than the Rmax figure on those linpack tests. These kinds of applications would tend to favor the x86 clusters on the top500 list more heavily, because in raw MIPS they are quite strong.