CASE STUDY Cray System “Blue Waters” Helps Researchers Develop a Defense Against Damaging Space Weather Organization Situation National Center for Supercomputing Applications Urbana-Champaign, IL www.ncsa.illinois.edu The sun may be the key to life-sustaining functions here on Earth. But it also poses a significant threat. The Earth is embedded in the Sun’s extended atmosphere and, as a result, the Earth and its technological systems are in constant threat from magnetic storms on the Sun. While most of these storms are not directed towards the Earth, when one does come our way the results can be devastating. About Blue Waters Housed at the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Urbana-Champaign and supported by the National Science Foundation, the Blue Waters supercomputer is one of the most powerful supercomputers in the world and the fastest system anywhere on a university campus. Scientists and engineers use the computing and data power of Blue Waters to tackle a wide range of challenging problems, from predicting the behavior of complex biological systems to simulating the evolution of the cosmos. The system is based on Cray® XE6™ and XK7™ technology and uses hundreds of thousands of computational cores to achieve over 13 petaflops of peak performance. System Overview •System: Cray XE/XK hybrid supercomputer •Cabinets: 288 •Peak Performance: 13 PF •System Memory: 1.5 PB •XE Compute Nodes: 22,640 AMD 6276 “Interlagos” processors •XK Compute Nodes: 4,224 NVIDIA® GK110 “Kepler” GPU accelerators •Interconnect: Gemini “The time we have on this machine is really precious and it’s limited, so we need to get the codes to run as efficiently as possible. Working with the [team] has been extremely helpful.” —Homa Karimabadi Space Physics Group Leader UCSD Cray Inc. 901 Fifth Avenue, Suite 1000 Seattle, WA 98164 Tel: 206.701.2000 Fax: 206.701.2500 www.cray.com At one time, solar storms had little detrimental effect on the Earth. But now that our daily lives depend on electronics, satellites and power grids, what’s happening on the surface of the Sun really starts to matter. A strong storm can disable high-voltage transformers, knock satellites out of orbit and cripple communications worldwide. The figures register in the billions of dollars when talking about losses in satellite technology that have been linked to space weather damage. The Sun consists of a hot, ionized gas called plasma with an embedded magnetic field. The more the plasma roils, the more it moves the magnetic field around, causing extreme tension. At some point the magnetic field lines snap, heating and accelerating the plasma in the process. This gives rise to solar flares and coronal mass ejections (CME). CMEs spew out billions of tons of matter and electromagnetic radiation into space. A flare can easily span 10 times the Earth’s diameter in size and release energy in the order of 160 million megatons of TNT equivalent. With the release of CMEs into the solar wind — the medium between the Sun and the Earth — a phenomenon known as space weather occurs. Mitigating the impact of solar storms is the focus for space physics group leader Homa Karimabadi from the University of California, San Diego and his research team. Challenge The solution — and the challenge — to coping with the effect of solar storms on the Earth and its technological systems is forecasting this space weather. A given CME can take anywhere from one to five days to reach the Earth, providing sufficient time to take evasive action. However, during solar maxima (a period of greatest solar activity in the 11 year solar cycle of the Sun), there can be over three solar storms per day. “You can’t just shut off the power grids, the electronics, and the satellites every time there is a storm. We really need to be able to judge the severity of a given storm, its impact, and its location of impact,” says Karimabadi. The challenge then becomes to develop accurate space weather forecasting models that can be put to use in real-time situations. While the Earth has a built-in defense called the magnetosphere that deflects most of the energetic particles and radiation coming from the Sun, a process called magnetic reconnection allows some solar wind to penetrate the planet’s protective shield. In order to forecast space weather, the research team must simulate and understand the physics behind this magnetic reconnection, a process that occurs on electron scales but has global consequences. Existing global magnetospheric codes can model the interaction of the solar wind with the Earth’s magnetosphere, says Karimabadi, but the ultimate goal is to run these codes in real time based on measured properties of a given CME heading toward the Earth and predict the geographical location and the severity of the impact. An added challenge is that the current models lack certain details, including the proper physics of magnetic reconnection essential for developing accurate forecast models. Thus much of his team’s effort is going toward developing models of magnetic reconnection that can be inserted into the global codes. The types of simulations required by this kind of research are some of the most challenging in terms of the data and memory requirements. ©2014 Cray Inc. All rights reserved. Cray is a registered trademark of Cray Inc. All other trademarks mentioned herein are the properties of their respective owners. 20140107_V1KJL CASE STUDY Page #2 Solution A petascale computing allocation received through the National Science Foundation enabled Karimabadi and his research team to prepare their codes for extreme-scale supercomputers and tap into the computing and data power of Blue Waters. “We wanted to know if we changed the parameters, such as system size, would the physics change,” Karimabadi says, “and if it did, could we develop scaling laws to extrapolate the results to real systems in nature.” The Cray® XE™/XK™ hybrid “Blue Waters” supercomputer made it possible for the team to push their simulations to the largest size possible on any supercomputer today. But even more, technical collaborations between the NCSA Blue Waters team, Cray, the San Diego Supercomputer Center and Los Alamos National Laboratory helped the team optimize their codes. One collaboration resulted in increased performance of their H3D code — a global code that treats electrons as fluids and ions as kinetic particles. Another resulted in improvements to VPIC — the kinetic code that models individual particles for the inclusion of electron physics — as a part of the sustained petascale performance (SPP) optimization effort. Prior to petascale computing, almost all of the global simulations were based on fluid models. But Blue Waters has given the team new options. With hundreds of thousands of computational cores, the system is enabling global simulations that include important ion kinetic physics using H3D code and local simulations that resolve both ion and electron kinetic physics using the code VPIC. The performance gains took some work, though. NCSA research programmer Kalyana Chadalavada implemented the use of more fused multiply-add (FMA) instructions, while Jim Kohn of Cray eliminated some redundant stores to temporary variables and reordered some independent Streaming SIMD Extensions (SSE) instructions for better resource utilization on the chip. Chadalavada says they were able to achieve an improvement of 12–18 percent by using more FMA instructions, as well as another 5 percent improvement by eliminating some redundancy and rearranging some parts of the code. “At scale, a significant portion of the run time is spent transferring particle data between processors. Changing the relevant communication routines by, for example, overlapping more communication with computation, could reduce the total simulation time considerably,” Chadalavada says. The holy grail of space plasmas and space physics would be to figure out how to model magnetic reconnection—which we now know depends heavily on and is strongly affected by electron kinetic effects—in a global code where it is not possible to resolve electron kinetic effects. Capturing the essence of magnetic reconnection in this way is the key to precise prediction of space weather and Karimabadi’s team is starting to see that breakthrough form in their future. In fact, their work with NCSA and Blue Waters has already produced a major revolution for their discipline says Karimabadi. In a fluid model, there aren’t any particles so there aren’t things bouncing off or deflecting. There is nothing to generate waves, so a lot of important information is lost. MODELING SPACE WEATHER: Magnetic field lines as visualized by the line integral convolution technique. The color is based on the magnitude of the magnetic field. The “closed” lines along the magnetopause current sheets are 3-D flux ropes. “There have long been speculations about how the electron and ion kinetic effects affect details of the reconnection process and its global consequences. Many of the ideas can now be tested and new ideas and questions formulated to make further progress,” Karimabadi says. “Now that we have been able to go beyond fluid models, which ignore details on small scales, we have uncovered new and unexpected effects and are finding ample evidence that physical processes occurring on small scales have global consequences.” Overall, the team has discovered new regimes of reconnection and scaling relations with system size and parameters. “It’s almost like you have always had really poor eyesight and someone gave you glasses for the first time. You start to see a lot of details that were completely absent in the previous simulations,” says Karimabadi. For more information on NCSA’s space weather work, visit www.ncsa.illinois.edu/news/story/so_much_detail. Content courtesy of National Center for Supercomputing Applications at the University of Illinois Urbana-Champaign