Raising standards for reproducibility and rigor in scientific research is a growing concern across all science, and is particularly relevant to the area of high-performance computing and the science it enables. Some unique questions pertain to this sector, like: What is the purpose of requiring code and data to be open, when other researchers don’t have access to the same or even comparable computers? What if the research findings specifically address computational performance, an attribute that is notoriously hard to replicate as you move to different hardware? When is achieving numerical reproducibility in finite-precision parallel computations a sensible demand?
The situation is further complicated in domains where software stacks combine complex numerical libraries and domain-specific code written over many years, and where computations run in leadership facilities via competitive allocations. The objective of the TCHPC Reproducibility Initiative is to lead a broad and deep conversation to advance the standards of simulation- and data-based science, and to work with the community to coordinate efforts in this important area as well as to document experiences and effective practices.
Manish Parashar, Distinguished Professor of Computer Science,
Rutgers, the State University of New Jersey