This is another guest blog we will be featuring from time to time. Here is Cliff Mass's part 2 blog of his concerns about the heart of weather forecasting, numerical weather prediction, or NWP. Here is the link to part 1. We will next post a response from other scientists working in this critical field of weather forcasting.-Bob Ryan
In my last blog on this subject, I provided objective evidence of how U.S. numerical weather prediction (NWP), and particularly our global prediction skill, lags between major international centers, such as the European Centre for Medium Range Weather Forecasting (ECMWF), the UKMET office, and the Canadian Meteorological Center (CMC). I mentioned briefly how the problem extends to high-resolution weather prediction over the U.S. and the use of ensemble (many model runs) weather prediction, both globally and over the U.S. Our nation is clearly number one in meteorological research and we certainly have the knowledge base to lead the world in numerical weather prediction, but for a number of reasons we are not. The cost of inferior weather prediction is huge: in lives lost, injuries sustained, and economic impacts unmitigated. Truly, a national embarrassment. And one we must change.
In this blog, I will describe in some detail one major roadblock in giving the U.S. state-of-the-art weather prediction: inadequate computer resources. This situation should clearly have been addressed years ago by leadership in the National Weather Service, NOAA, and the Dept of Commerce, but has not, and I am convinced will not without outside pressure. It is time for the user community and our congressional representatives to intervene. To quote Samuel L. Jackson, enough is enough. (click on image to watch him say that famous line)
Enough is Enough! The U.S. Needs Better NWP
- NCEP (U.S) Computer
The European Centre has a newer IBM machine with 8192, much faster, processors that gets 182 terraflops (yes, over twice as fast and with far fewer tasks to do). The UKMET office, serving a far, far smaller country, has two newer IBM machines, each with 7680 processors for 175 teraflops per machine. Here is a figure, produced at NCEP that compares the relative computer power of NCEP's machine with the European Centre's. The shading indicates computational activity and the x-axis for each represents a 24-h period. The relative heights allows you to compare computer resources. Not only does the ECMWF have much more computer power, but they are more efficient in using it...packing useful computations into every available minute.
The European Centre has a newer IBM machine with 8192, much faster, processors that gets 182 terraflops (yes, over twice as fast and with far fewer tasks to do).
- Courtesy of Bill Lapenta, EMC
Recently, NCEP had a request for proposals for a replacement computer system. You may not believe this, but the specifications were ONLY for a system at least equal to the one that have. A report in a computer magazine suggests that perhaps this new system (IBM got the contract) might be slightly less powerful (around 150 terraflops) than one of the UKMET office systems...but that is not known at this point.
The Canadians? They have TWO machines like the European Centre's!
So what kind of system does NCEP require to serve the nation in a reasonable way?
To start, we need to double the resolution of our global model to bring it into line with ECMWF (they are now 15 km global). Such resolution allows the global model to model regional features (such as our mountains). Doubling horizontal resolution requires 8 times more computer power. We need to use better physics (description of things like cloud processes and radiation). Double again. And we need better data assimilation (better use of observations to provide an improved starting point for the model). Double once more. So we need 32 times more computer power for the high-resolution global runs to allow us to catch up with ECMWF. Furthermore, we must do the same thing for the ensembles (running many lower resolution global simulations to get probabilistic information). 32 times more computer resources for that (we can use some of the gaps in the schedule of the high resolution runs to fit some of this in...that is what ECMWF does). There are some potential ways NCEP can work more efficiently as well. Right now NCEP runs our global model out to 384 hours four times a day (every six hours). To many of us this seems excessive, perhaps the longest periods (180hr plus) could be done twice a day. So lets begin with a computer 32 times faster that the current one.
Many workshops and meteorological meetings (such as one on improvements in model physics that was held at NCEP last summer---I was the chair) have made a very strong case that the U.S. requires an ensemble prediction system that runs at 4-km horizontal resolution. The current national ensemble system has a horizontal resolution about 32 km...and NWS plans to get to about 20 km in a few years...both are inadequate. Here is an example of the ensemble output (mean of the ensemble members) for the NWS and UW (4km) ensemble systems: the difference is huge--the NWS system does not even get close to modeling the impacts of the mountains. It is similarly unable to simulate large convective systems.
- Current NWS( NCEP) "high resolution" ensembles (32 km)
- 4 km ensemble mean from UW system
Let me make one thing clear. Probabilistic prediction based on ensemble forecasts and reforecasting (running models back for years to get statistics of performance) is the future of weather prediction. The days of giving a single number for say temperature at day 5 are over. We need to let people know about uncertainty and probabilities. The NWS needs a massive increase of computer power to do this. It lacks this computer power now and does not seem destined to get it soon.
A real champion within NOAA of the need for more computer power is Tom Hamill, an expert on data assimilation and model post-processing. He and colleagues have put together a compelling case for more NWS computer resources for NWP. Read it here.
Back-of-the-envelope calculations indicates that a good first step-- 4km national ensembles--would require about 20,000 processors to do so in a timely manner--but it would revolutionize weather prediction in the U.S., including forecasting convection and in mountainous areas. This high-resolution ensemble effort would meld with data assimilation over the long-term.
And then there is running super-high resolution numerical weather prediction to get fine-scale details right. Here in the NW my group runs a 1.3 km horizontal resolution forecast out twice a day for 48h. Such capability is needed for the entire country. It does not exist now due to inadequate computer resources.
The bottom line is that the NWS numerical modeling effort needs a huge increase of computer power to serve the needs of the country--and the potential impacts would be transformative. We could go from having a third-place effort, which is slipping back into the pack, to a world leader. Furthermore, the added computer power will finally allow NOAA to complete Observing System Simulation Experiments (OSSEs) and Observing System Experiments (OSEs) to make rational decisions about acquisitions of very expensive satellite systems.
The fact that this is barely done today is really amazing and a potential waste of hundreds of millions of dollars on unnecessary satellite systems. But do to so will require a major jump in computational power, a jump our nation can easily afford. I would suggest that NWS's EMC should begin by securing at least a 100,000 processor machine, and down the road something considerably larger. Keep in mind my department has about 1000 processors in our computational clusters, so this is not as large as you think.
- For a country with several billion-dollar weather disasters a year, investment in reasonable computer resrouces for NWP is obvious.
The cost? Well, I asked Art Mann of Silicon Mechanics (a really wonderful local vendor of computer clusters) to give me rough quote: using fast AMD chips, you could have such a 100K core machine for 11 million dollars. (this is without any discount!) OK, this is the U.S. government and they like expensive, heavy metal machines....lets go for 25 million dollars. The National Center for Atmospheric Research (NCAR) is getting a new machine with around 75,000 processors and the cost will be around 25-35 million dollars. NCEP will want two machines, so lets budget 60 million dollars. We spend this much money on a single jet fighter, but we can't invest this amount to greatly improve forecasts and public safety in the U.S.? We have machines far larger than this for breaking codes, doing simulations of thermonuclear explosions, and simulating climate change.
Yes, a lot of money, but I suspect the cost of the machine would be paid back in a few months from improved forecasts. Last year we had quite a few (over ten) billion-dollar storms....imagine the benefits of forecasting even a few of them better. Or the benefits to the wind energy and utility industries, or U.S. aviation, of even modestly improved forecasts. And there is no doubt such computer resources would improve weather prediction. The list of benefits is nearly endless. Recent estimates suggest that normal weather events cost the U.S. economy nearly 1/2 trillion dollars a year. Add to that hurricanes, tornadoes, floods, and other extreme weather. The business case is there.
As someone with an insider's view of the process, it is clear to me that the current players are not going to move effectively without some external pressure. In fact, the budgetary pressure on the NWS is very intense right now and they are cutting away muscle and bone at this point (like reducing IT staff in the forecast offices by over 120 people and cutting back on extramural research). I believe it is time for weather sensitive industries and local government, together with t he general public, to let NOAA management and our congressional representatives know that this acute problem needs to be addressed and addressed soon. We are acquiring huge computer resources for climate simulations, but only a small fraction of that for weather prediction...which can clearly save lives and help the economy. Enough is enough.