Slashdot Mirror
Supercomputing Upgrade Produces High-Resolution Storm Forecasts
dcblogs writes A supercomputer upgrade is paying off for the U.S. National Weather Service, with new high-resolution models that will offer better insight into severe weather. This improvement in modeling detail is a result of a supercomputer upgrade. The National Oceanic and Atmospheric Administration, which runs the weather service, put into production two new IBM supercomputers, each 213 teraflops, running Linux on Intel processors. These systems replaced 74-teraflop, four-year old systems. More computing power means systems can run more mathematics, and increase the resolution or detail on the maps from 8 miles to 2 miles.
Now they can be wrong in hi-def!
How can I believe you when you tell me what I don't want to hear?
I was at a supercomputing conference back in the 90's. There were wonderful reports on doubling the resolution of the grid and so on. Advances in the scale are all good.
The questions are
a) with the increase in detail of the simulations have we converged on a solution. That is do solutions at scale N and 10N match. If they do then the resolution and model are aligned for accuracy in the solution.
b) do the simulations agree with reality.
If a) and not b) then there is something wrong with the model that is not related to compute power or problem resolution, and no amount of compute power will fix it.
Better resolution is good, but with each improvement in the system, the input data also needs to be improved and remeasured.
Ultimately the ground features need to be modelled in greater detail to match the increased resolution of the grid.
Which comes own to knowing where each tree/building and similar sized static feature is and how this affects the model.
However, as the grid increases it should not need to know where the butterflies are .
HRRR going into production is not directly a result of the supercomputer upgrade. It is a new model that has been in testing for two years, but has just recently been officially "released". While the supercomputer upgrade is great for the increased runs and spatial resolution, it is not directly related to the HRRR being a better model. It will, however, be much more useful when the GFS model is upgraded later this year to significantly increased horizontal and vertical resolutions.
Why is the European Weather Forecast more accurate than the American/USA Weather Forecast model? Also, Intel takes short cuts with their CPU. Intel CPU truncates the math and maybe taking other shortcuts.
You really are pretty clueless about this computing stuff.
Please stop posting about things you only have a vague understanding of. It's not a good look.
Contributing to global warming, no doubt.
Thanks guys for buying the iSuper. 4x more powerful, 4x the resolution. Come back in 4 years for the retina resolution.
I'm a computer engineer, not a meteorologist, but I've worked with them off and on for about eight years now. One of the most common models for research use is "Weather Research and Forecasting Model" (WRF, pronounced like the dude from ST:TNG). There are several versions in use, so caveats are in order, but in general WRF can produce really good results at a 1.6KM grid for 48 hours in the future. I was given the impression that coarser grids are the route to happiness for longer period forecasts.
WRF will accept about as much or as little of an initializer as you want to give it. Between NEXRAD radar observations, ground met stations all over the place, two hundred or so balloon launches per day, satellite water vapor estimates, and a cooperative agreement with airlines to download in-flight met conditions (after landing, natch), there's gobs of data available.
The National Weather Service wants to run new models side-by-side with older models and then back check the daylights out of them, so we can expect the regular forecast products to improve dramatically over the next (very) few years.
Well, congrats to NOAA for the milestone...but isn't a peta-scale system a thing of the past for something as computationally intensive and critical as weather prediction? The prediction model seems to have no scalability issues: "The Hrrr model produces output from the model every 15 minutes versus the previous hourly rate,..", meaning 4x speedup for a 4x upgrade in peak performance. Why didn't NOAA go for a peta-scale system straightaway?
The upgraded systems are already placed way down the list of world's fastest 500: http://top500.org/system/17783....
The article says a peta-scale system is under way but I'm afraid by the time it starts ringing its bells, we'll all have entered exa-scale. Is NOAA always going to be catching up?
China has a much bigger computer with over 3 million cores.
Also, aren't the two "new" supercomputers 2 years old now?
I am in an area in which hurricanes are common. After a storm we see wind reports that indicate speeds like 120 mph or 140 mph. But people who live here see I beams that are quite large bent like noodles and we all know that it took much higher winds to bend those beams. Apparently micro bursts are the culprit and these tornados must have immense strength. We do not normally hear where these officially recognized bursts have occured but it does explain why some homes are simply inconvenienced a bit while other homes are erased from the face of the Earth. I would love to know the wind speeds that happen in these bursts and I'm betting on at least 300 mph. The girders that I mention were bare and holding nothing. Just the wind on the girders bent them down at right angles.
This thing is going to put all of the local weathermen out of jobs. I mean how can they predict the weather better than this thing? ;-)
Keep the Classic Slashdot.
July 2013.
Has the resolution and reliability of initial data points improved as well? Or are we just doing a finer interpolation of model output with same data input?
There have been advances in assimilating radar and satellite data into models, which can provide better resolution than data from other observation platforms such as profilers, raobs, and surface observing stations. This reduces the time needed for models to spin up.
The real advantage from improving the resolution is the ability to explicitly represent processes in a way that couldn't be done with coarser models. This is especially true in regard to convective processes. A thunderstorm updraft might be a few kilometers across. In order to explicitly represent this at all in a model, the horizontal grid spacing should be 4 km or better. The GFS isn't a gridded model but a spectral model; however, it's able to resolve features down to around 27 km. It's not close to resolving actual thunderstorms. The HRRR is initialized off the RAP, which has a 13 km horizontal grid spacing. The NAM is a mesoscale model with a horizontal grid spacing around 12 km. I believe the NAM is around 12 km, though there are some higher resolution versions.
A model with a coarser resolution must parameterize the convection, rather than explicitly simulating it in the model. It's far less accurate than explicitly simulating the convection. The cumulus parameterization schemes exist to remove instability in models, which is essential for their performance, but don't necessary do a good job in actually forecasting the convection. Furthermore, it's not possible for the forecaster to see the storm mode. For example, will the storms be supercells (greater risk of large hail and tornadoes) or a squall line (greater risk of wind damage)? That's an important question that the HRRR can answer far better than the NAM or GFS.
Also, the Hurricane WRF (HWRF) is now run with an inner nest of 3 km. Now, we don't get nearly good enough observations to have a good estimate of the initial state of the atmosphere with a horizontal grid spacing at 3 km. But new observations are assimilated using the previous model forecast as a first guess. It does an adequate job of predicting the initial state of the system. Because the model is run at 3 km, however, it explicitly resolves the convection at the inner core of a hurricane. This is very useful in predicting the intensity of the storm. The GFS, at roughly 27 km, just isn't suitable for forecasting hurricane intensity, but the HWRF can.
It's not necessarily that the observations have improved, though models assimilate more high-resolution observations than ever. However, improving the resolution allows the models to resolve processes that previously needed to be parameterized. Explicitly resolving convection definitely improves the accuracy of forecasts involving thunderstorms and hurricanes.
I posted, so can someone else mod the parent up? tnx.
while (1) {
prediction = PredictWeather();
if (prediction == true) {
AskForMoreGrantFunding();
BlaimGlobalWarming();
} else {
BlaimGlobalWarming();
AskForMoreGrantFunding();
}
}
Of course, you can get estimates for water vapour from IR satellite measurements. I saw this done also in the 1980s. At the time I didn't understand all the math used to do this, but remember that it involved taking IR emissions over several wavelength bands, and somehow combining these to infer the water vapour content at various heights in the atmosphere, under each pixel. These satellites certainly have 2-mile spatial resolution, but the problem I see there is that the polar-orbiting satellites that provide this information pass over any spot on earth about 4 times per day, so the temporal resolution is as low as the balloons'.
Finally, data from airlines is going to be largely restricted to heights at cruising altitude, so you're missing a large cross section of the atmosphere there.
And don't get me started on weather observations over the ocean, where there are very few ground stations or balloons.
The issue is that the Navier- Stokes equations being solved in weather forecasting are very sensitive to initial conditions, so it's really crucial to get the data right to set up the calculations. Sitting in my armchair, I remain a bit skeptical that we will ever be able to get the true initial conditions.
Despite all this I'm always impressed that the NWS manages to get pretty decent one-week forecasts out, despite the impossible task they face. There must be some deep voodoo in those numerical models!
This was an interesting article. However as UW's Cliff Mass has previously pointed out (And, today, he privately confirmed is still the case), NOAA is sitting on already-approved funds to purchase a weather modeling computer that's seen as a potential "game changer" for US climate modeling.
Over a year ago Congress approved the purchase of a computer that's roughly an order of magnitude more powerful than the pair mentioned in this article - but, because NOAA has a contract with IBM and IBM recently sold their server business to Lenovo, NOAA has been sitting on their hands regarding approval of the purchase of such a computer from a Chinese company.
So while the improvements mentioned in the article are better than nothing... in truth we should be a significant step beyond that by now.
#DeleteChrome