This is a contributed blog post from University La Sapienza (Rome, Italy) PhD Student, Alessandra Mastrobuono Battisti, and professor of astronomy and astrophysics, Roberto Capuzzo-Dolcetta.
Seems like humans have always had a fascination with the stars—looking up into the sky and wondering about galaxies far away. In fact, astronomy, or the study of the universe and stars, is perhaps one of the oldest scientific disciplines.
Today with the advent of powerful computers, astronomers can use complex mathematical models to study how stars interact. They can see the formation and collisions of star clusters and galaxies, where black holes exist.
Looking at the video below, you can see how a GPU simulation, created over 5 hours, shows the radial fall of a star cluster toward a massive black hole in the center of our galaxy. This black hole has a mass 10 times greater than the star cluster
At the University of Rome, we developed a new software application called NBSymple that takes advantage of a hybrid combination of CPUs and GPUs. It uses the OpenMP directives model along with the CUDA parallel programming model for GPUs. For the hard core technies out there, we used a high precision, symplectic, algorithm for time integration, which guarantees energy conservation.
We compared our software with various processor technologies, using both single precision and double precision floating point arithmetic. The results were dramatic — just one Tesla C1060 GPU ran 330 times faster than 2 Quad Core Intel Xeon 2.00 GHz CPUs for a star system with N=7680 stars. When we added the 2nd Tesla C1060 GPU, the application ran two times as fast.
It took 2 Tesla GPUs 1 minute to evaluate all the forces among N=1,536,000 stars in a stellar Globular Cluster; this means that in less than 9 days of simulation we can model a full revolution of the Globular Cluster around the center of the Milky Way.
The “real time” revolution of a cluster in the Milky Way takes about 250 million years to be accomplished.
This kind of order of magnitude improvement in simulation enables our research group to study much more complex interactions in the dynamics of self-gravitating systems.