by Benjamin Recchie, AB’03
Cosmologists used to have it easy. Not that cosmology was ever easy, of course. But when there weren’t that many sources of data to choose from, a cosmologist didn’t have to worry about how to make a model that fit them all. In the last decade, though, the number of data sets available to researchers has risen greatly. Now, they need to fit data from the Wilkinson Microwave Anisotropy Probe (WMAP), the Sloan Digital Sky Survey (SDSS), the South Pole Telescope (SPT), and many others.
Each one adds to our understanding of the universe, but each was created with different instruments with different strengths and limitations. A model that fits one data set nicely might not match up well with observations from another project. Furthermore, modifying a modeling program to account for new theories is difficult. To top it off, research collaborators around the world might not use the exact same type of computer to compile and run the programs; the same program might produce slightly different results if run on two dissimilar machines.
This is the problem the researchers of the Dark Energy Survey (DES) faced when they were planning for the analysis of their data. DES aims to better understand the accelerating expansion of the universe and uncover the nature of dark energy, the poorly understood substance that makes up a majority of the cosmos, by using observations from a single telescope to attack the problem of dark energy from four different angles (called “probes”). Each probe measures different quantities and has its own errors and biases, and multiple models might be needed to explain all the observations. In addition, the international collaboration—more than 120 scientists from 23 institutions across the world—wanted to make sure their analysis program gives them the same results regardless of what kind of computer it runs on.
The collaboration includes Scott Dodelson, a scientist at Fermilab and professor of astronomy and astrophysics and the Kavli Institute for Cosmological Physics (KICP) at the University of Chicago. Dodelson heads a working group for DES that tried to find a way to provide a consistent and flexible way to analyze their result, a cosmological modeling program that could account for differences in the probes in their own data set, as well as use data from other experiments. The group, which also includes Joe Zuntz and Sarah Bridle at the University of Manchester, Marc Paterno, Jim Kowalkowski, and Saba Sehrish at Fermilab, and Elise Jennings and Alessandro Manzotti at KICP, turned to the Research Computing Center (RCC) for assistance in designing such a program. In turn, RCC provided them with the services of Douglas Rudd, PhD’07, who had earned his doctorate in astrophysics from UChicago and understood the nuances of both cosmology and computer programming.
The collaboration came up with a program called the Cosmological Survey Inference System, or CosmoSIS for short. They designed it to be modular, with different functions compartmentalized so that researchers could insert new properties or aspects of the model without rewriting the whole program. “The modularity is the key,” says Dodelson. “It enables anyone to stick in his or her own module while making use of much code and many results from other experiments with minimal overhead.”
Another way the team makes sure CosmoSIS is as useful as possible is by making sure “the compiler [gcc], library versions, and the way the user installed the program don't vary in a way that can produce incorrect results,” explains Rudd. To combat that, CosmoSIS is distributed with all of its dependencies, including its own numerical libraries. Rudd says they’re also supporting a limited number of systems that they've tested CosmoSIS on to ensure the packages produce consistent results. He also notes that since it's also an open-source program, he anticipates that its users will also contribute their own modules
CosmoSIS is already available to anyone, though Dodelson says they’re still working on improving the program. The developers are curating a library of new modules that researchers can download and plug in to CosmoSIS—and encouraging others outside DES to contribute their own. So while DES is two years into its projected five-year run of taking data, CosmoSIS will likely long outlive it.