The advance of scientific computing rapidly removes obstacles that appear impossible to solving numerical problems
(c) 12 May 2012 Fernando Caracena
Before 2003, several of us at the Forecast Systems Labs (FSL), in Boulder Colorado, discussed a phenomenon that foreshadowed heavy rain or severe weather: long, narrow slots of dry air in the upper troposphere appeared in the water vapor satellite imagery in advance of large storms that brought either heavy rain or severe weather over state-wide areas of the United States. Calculations of potential vorticity (PV) contours overlaid over satellite images indicated that the dry slots represented elongated streaks of potential vorticity, which in Europe were called PV Streamers. Either heavy rain or severe weather happened at the intersection of a PV Streamer and a frontal boundary near the surface that was being overrun by a flow of moist air. Or, instead of a frontal boundary, the moist airflow could be up the side of a mountain.
A new massively parallel computer was under construction at the time at FSL, which the modelers were invited to try. A friend, Dr. Adrian Marroquin from Colombia, who was a mesoscale modeler, agreed to study the phenomenon by modeling some of the cases, which we had analyzed and documented. At the time the new parallel computer had only 128 nodes (later the number of CPUs was doubled and doubled again), of which we were allowed to use only 12 in our modeling efforts. Even so, we were fairly successful in modeling one of the cases, the results of which we published in a European Journal[1], there being hardly any interest in this phenomenon in the United States at the time.
We could have continued to model the role of PV Streamers in organizing the large scale structure of storms, but our efforts were thwarted by the laboratory administration when Adrian was denied access to FSL's parallel computer network. Now computer technology and science has advanced to the point that we could do the PV streamer simulations on my desktop without asking for anybody's permission.
I recently upgraded my desktop computer by adding an NVIDA GeForce GTX480 video card that contains 480 GPUs. I also upgraded my Linux operating system from Kubuntu 10.04 to Kubuntu 12.04. In the process, I gained support for using the GPU processors in Python through the Fenics system and Dolfin libraries. I tried an example, which solved Poisson's equation on a circular finite element mesh of a radius of 40 radial grid points (over 5,000 nodes) in less than a minute using FENICS software and Dolfin libraries. Most of the computation time involved python's setting up the parameters used by the GPU-related software.
Python is an excellent language for scientific computing
Python, which is a scripting language, looks to me to be a very handy computational language, especially since it has both good and free GPU support through the Dolfin libraries and the FENICS project. Several programming languages have support for my NVIDA 480 GT card using these processors in parallel processing. Some of them are scripting languages. But I find that python gives me the ease of programming through scripting, and yet allows me to use massively parallel processing to do fast number crunching, which is provided by free libraries.
Thanks to the shortening doubling times for progress in computer design, I now have a personal computer that has a greater capability than the Boulder Labs' parallel computer had in the 1990s.
What this experience has taught me is that it is now feasible to wait for the advance of science and technology to clear away hurdles (both technical and administrative) that seem impossible in solving a personally posed scientific problems. In the near future, large laboratories may have much less to offer highly motivated scientists, except perhaps to provide a place where groups of scientists can gather and do their own thing, and of course, funding. They will not have much say in determining what a scientist can or cannot do in the way of research when he is at home. But even here, the connectivity of the Internet may offer still more. Science could be done in the future globally from home without a big supervisory structure, which leaves only the requirement of funding of scientific projects.
1. Caracena, F., Marroquin, A., and Tollerud, E.: A PV-streamer’s role
in a succession of heavy rain-producing MCSs over the central
United States, Phys. Chem. Earth, Part B-Hydrology Oceans and
Atmos., 26(9), 643–648, 2001.
2 Responses to GPUs and the Internet facilitate doing science globally from home