Boulder, Colorado - June 26, 2018:
Tech-X CEO and Colorado Prof. PhysicsJohn Cary Presents Plenary Talk at 2018 ICOPS
On June 25, 2018, Tech-X CEO and Colorado Prof. Physics John Cary, 2016 recipient of the IEEE NPSS Charles K. Birdsall Award for Contributions to Computational Nuclear and Plasma Sciences, presented a plenary talk, "Evolution of Computational Physics", at the 2018 International Conference on Plasma Science.
The abstract of the talk is reprinted here:
THE EVOLUTION OF COMPUTATIONAL SCIENCE AND ENGINEERING
John R. Cary
Tech-X Corporation, Boulder, CO 80303
Department of Physics, U. Colorado,
Boulder, CO 80309-0390
The evolution of computational physics over the last 30 years has been frighteningly fast, and the change has affected every aspect of the trade, including education, hardware, operating systems, languages, methodologies, standards, available libraries, information sharing, teaming, management, and specialization. This talk will concentrate on those changes to computational plasma and beam physics and engineering that result from the increasing specialization of the field, for which computationalists now typically concentrate on the computer science aspects, on the applied math issues. on the science implementation, or on science discovery.
Several computational discoveries are described. Laser-plasma acceleration has come to depend heavily on simulation to identify the physical mechanisms for variation of the beam emittance in the direction of and transverse to laser polarization. Many ideas are also first tested by computation, which can be done more rapidly than experiments can be reconfigured. Similarly, propagation of radiation in magnetized plasmas has been analyzed computationally; the computations show that propagation of electromagnetic energy into plasma is difficult for frequencies somewhat above the gyrofrequency because of the many nonlinear decay channels.
Improvements include algorithms that preserve the properties of the underlying system, are efficient on distributed memory computers, and take advantage of vector parallelism of GPUs and Advanced Vector Instructions. Regarding the former, we discuss structure-preserving algorithms that maintain properties of the underlying equations (that is, volume preservation) and hence allow computation of sensible equilibria even though the solutions are only second-order accurate. For numerical efficiency, new multi-step algorithms allow better use of cache and lead to factors of several in speedup.
Computation has become sufficiently complex that practitioners must have a workflow that allows problem setup with an easy-to-use GUI, transition of the problem to a supercomputer, and capture of a provenance to replicate the results. A preferred workflow from GUI setup to supercomputer usage and data analysis will be described.
Finally, the rapid evolution has affected education, as it is no longer possible, within the graduate school training period, for a student to learn everything from the impact of CPU/GPU hierarchy on implementations through how to set up simulations to answer questions. To this end, the University of Colorado now offers multiple courses, including applied math courses on parallel algorithms to physics courses on how to make effective use of simulations. The latter will be described.