XSEDE14 Workshop Wrestles with Reproducibility
Imagine that you are trying to create a new sauce for a special dish, or the perfect adhesive for a new aircraft, or you’re flying a helicopter looking for victims of a natural disaster — and you succeed at each of these. This is wonderful news for your dinner guests, or the company that will use the new adhesive, and especially for the victims of the natural disaster. But the question is — Could you do it again and get the same results? Or, did you just get lucky the first time?
At the XSEDE14 conference in Atlanta, a roomful of computational veterans from inside and outside the NSF Extreme Science and Engineering Discovery Environment (XSEDE) participated in a full-day workshop on the topic of reproducibility, and clearly, there is a lot at stake.
“There is a growing awareness in the computational research community that this question of ‘can we do it again’ is becoming important for us in new ways, and the stakes are high — computational research is helping to save lives, answering policy questions, and making an impact on the world,” said Doug James, an HPC researcher at the Texas Advanced Computing Center, in his opening remarks for the workshop.
People have been thinking about reproducibility for a long time – it is one thing to reproduce a small scale lab experiment, or a computation on your desktop, but it is an entirely different matter to reproduce something that the Hubble Space Telescope did over five years at the cost of hundreds of millions of dollars, for example.
So, what is reproducibility? One working definition might resemble this: the ability to repeat an experiment to the degree necessary to assess the correctness and importance of the results. Practices that promote reproducibility include anything that makes a researcher more organized, provides a better audit trail, allows a researcher to track source code, and to know what data sources were used.
Victoria Stodden of Columbia University, who led a roundtable on the topic of reproducibility in 2009 and an ICERM workshop on Reproducibility in Computational and Experimental Mathematics in 2012, gave the keynote address at the XSEDE14 workshop. She raised the issue of a credibility crisis.
“Reproducibility has hit the popular press over the last several months,” Stodden said, citing recent coverage by The Economist (October 2013) and editorials in Nature and Science. Issues around the importance of reproducibility were catalyzed by the clinical trials scandal at Duke University in computational genomics where mistakes in the research were uncovered in 2010 in The Cancer Letter.
“This really goes to the heart of how important reproducibility issues are, and how we need to reconstruct the pipeline of thinking, reasoning and observation that a scientist does, but for the computational aspects, too, where many of these decisions are being manifest.”
Stodden also touched on separate discussions going on regarding different aspects of reproducibility such as statistical reproducibility, which questions the research decisions about the statistics and data analysis, and empirical reproducibility, which focuses on the reporting standards for the physical experiment, but does not focus on the computational steps.
Everyone in the room agreed that computational research is now in a position where complexity and mission criticality take on new import, and the community needs to develop confidence in the results of that research. But what should our priorities be? Training? Better tools? New steps in proposals and submissions?
NCSA Director Ed Seidel shared his view that there are three levels where things have to happen to get momentum moving in right direction: 1) campus level; 2) national level; and 3) publisher level.
Seidel said that local campuses have to think about how they can begin to support local data services, not just repositories, so there is a local structure. “This is a policy issue that vice chancellors for research and provosts need to take seriously…and there are organizations in place like Internet2 and Educause that span the research universities across the country that can help,” Seidel said. “It’s important to frame it not just as data but more around reproducibility; scope the problem beyond data and the data infrastructure.”
In addition, Seidel cited the XSEDE initiative as being a good organization for aiding the reproducibility process. XSEDE was instrumental in starting the National Data Service Consortium, aimed at organizing a number of individual efforts for data services around tools to create data collections to get Digital Object Identifiers or ‘DOIs’ associated with them and to provide linking services to publishers. While typically thought of as pointers to data collections, DOIs can also attach to code. This is a crucial part of reproducibility.
Professional societies and journals can play a part as well. Many are starting to require links to the data referenced in a publication. But reproducible practices must start in the research group.
Lorena Barba of George Washington University and a leading advocate of reproducible science said, “Conducting research reproducibly doesn’t mean someone else will reproduce the results, but that you are doing it as if someone would do this. By providing full documentation, access to input data and source code, the community will have confidence in your results and will label them as reproducible even if they are, in fact, not reproduced.”
Many other people added to the conversation including Mark Fahey of the National Institute of Computational Sciences. According to Fahey, the centers need to step up and take some responsibility for providing documentation about how users build and run their codes. Fahey said, “Centers can automatically collect information for each code built and each run of the code, and this information can be made available back to the researcher for publications if desired. There are already two prototypes (ALTD and Lariat) at a variety of computing centers around the world that collect a good portion of this information, and a new improved infrastructure is in development called XALT funded by NSF.”
At the outset of the workshop, the group committed to a key deliverable: recommendations in the form of priorities and initiatives for organizations and communities.
“It’s been implicit that ‘Of course, this is what people do, system administrators and researchers check to ensure that codes gets the same results after systems upgrades and when porting to new platforms’ but reproducibility has never been a formal enterprise,” said Nancy Wilkins-Diehr of the San Diego Supercomputer Center, who summarized the workshop and helped facilitate suggestions for moving forward.
“This is a good time to do this. Computational science is a respected contributor of the scientific knowledge base. Important decisions are now based on simulation. While this is gratifying, it has very real implications for our responsibilities as well,” she said.
The participants intend to move forward with humility, however. “The vision for the recommendations is to honor the reality of a diverse set of viewpoints and include ideas that might be outside of the box,” James concluded. Everyone agrees that there is a need to promote confidence-building tools and methodologies that do not adversely affect performance.
Recommendations will be ready in the September 2014 timeframe — please refer to xsede.org/reproducibility to read them. In addition, you can send comments and suggestions to email@example.com. The Help Desk will send any and all inquiries to the XSEDE team working on this initiative.