Cutting-Edge Physics Computing Is No Game

May 5, 2015

Ever resourceful, physicists, including several KIPAC scientists, have been using the specialized processors in computer graphics display cards to speed up some of the calculations that arise in data analysis. In the coming era of large astronomical surveys for weak lensing constraints on dark energy, such speed will be essential.

Example of the mass distribution (colors) inferred from analyzing the correlated subtle distorted shapes of galaxies at every point. The correlated distortion patterns are represented by the short lines.

Graphics processing units (GPUs) are special computing modules developed for the fast rendering of graphics, creating realistic texture and movement, with particular applications in computer gaming. They contain hundreds of processor cores and are designed to perform thousands of arithmetic calculations in parallel. General purpose computing on GPUs (GPGPU) is the computing strategy where the GPU works together with the computer's CPU (the more traditionally used processor core for calculations) taking advantage of the fast on-chip memory and parallel calculating capabilities of the GPU. The sequential part of an application can be run on the CPU, and the GPU handles the computationally intensive calculations in parallel. There are many applications of GPGPU in scientific computing, ranging from simulations of protein folding to calculations in quantum chromodynamics, and from weather modelling to astronomical simulations involving billions of individual particles.
 
GPGPU can be used to solve large-scale computational challenges in the era of "big data", such as the upcomming 7 petabytes (7 million billion bytes) per year that LSST will accumulate, but can also be used to speed up a scientist's everyday analysis work.  Most modern laptop and desktop computers contain a GPU card, which can be harnessed to accelerate calculations. KIPAC postdocs Deborah Bard and Mark Allen, along with Stanford Particle Physics postdoc Matt Bellis, have been exploring how to apply this technology to speed up their calculations, using publicly available GPU interfaces.  For example, a simple likelihood fit of a Gaussian function to a dataset of ten million data points is faster by a factor of 8 when calculated using GPGPU with an ordinary laptop's off-the-shelf GPU.
 
As an example of the potential benefits of GPGPU, weak lensing analysis studies the large-scale structure in the Universe - and the evolution of dark energy over cosmic time - through the small gravitational distortions of the light from distant galaxies as it passes near concentrations of matter. Detection of this 'cosmic shear' relies on correlating individual observed galaxy shapes on different distance scales. Future large surveys (such as LSST) expect to measure the shapes of billions of galaxies, and calculating the distance between each pair of galaxies is a daunting task.  Bard and Allen found that making use of the GPU to calculate many separation distances in parallel can speed up the calculation by a factor of 11 for a catalogue of ten million galaxies.  This speed up factor will scale with number of galaxies.
 
Another area where Bard and colleagues have shown GPUs to be beneficial is the hunt for mass peaks in the weak lensing maps - using the subtle changes in observed galaxy shapes to identify large foreground concentrations of dark matter in the Universe. This involves summing the contributions to the inferred unseen mass over tens of thousands of surrounding galaxies for each point on the sky, a task which can be sped up by a factor of 2 by calculating all the contributions in parallel on the GPU. In analyzing the weak lensing data from future large surveys such as LSST, where on the order of a billion galaxies will be considered to achieve unprecedented weak lensing constraints on the growth of structure and evolution of dark matter in the Universe, such computational techniques will be indispensible.
 
Other KIPAC scientists who have utilized GPGPU to enhance computing include the groups of Professors Tom Abel and Risa Wechsler, who perform extensive many-body cosmological simulations, where the tracking of the interactions between billions of simulated particles means that speed is at a premium.  A paper from Abel's group describing their use of GPGPU in this context is one of the most downloaded papers in the journal New Astronomy (New Astron, 2010, 15, 581).
Science Contact:
Deborah Bard
KIPAC
djbard@slac.stanford.edu
 
Tidbit Author: Deborah Bard and Jack Singal