It takes a whole community to build a sustainable ecosystem consisting of various independently developed software products that work well together. Providing a set of guidelines is an essential ingredient. This article discusses community policies for long-term development of a sustainable software ecosystem for scalable high-performance libraries and applications.
Building a software community
The design of the Extreme-scale Scientific Software Development Kit (xSDK, see https://xsdk.info) has been driven by the vision to provide a collection of related but highly diverse software products that enable and support the rapid development of large, well-performing application codes. The realization that software development is a community process was the initial thought that guided the xSDK project. This approach may be partially attributed to the interdisciplinary nature of computational science in general and to the xSDK's founding team in particular. Furthermore, the race toward extreme-scale computing became a true catalyst of the xSDK effort, with its focus on achieving a productive and sustainable software environment. We contrast this endeavor with moonshot-like projects that achieve a narrowly defined goal, where any benefits beyond the specific target are purely coincidental. By bringing together community-building qualities and software-development practices, the xSDK is working to achieve longevity by following a cohesive set of principled guidelines that now benefit the broad efforts to make exascale computing a reality. The founding packages that conformed to the initial set of community policies were few. Over time, however, the number has grown and is now approaching 20, with new packages joining every year (see the figure above).
Basic design of community policies
At the onset of the xSDK project, it became clear that to achieve the seamless coordination required to build a variety of software libraries---that are actually developed independently at different institutions---required a set of guidelines. To this end, the founding representatives of the xSDK member packages put together an initial draft of xSDK community policies. This draft was a result of lengthy discussions among software experts who all shared the goal of overcoming incompatibilities among their respective software packages that hindered building and using the xSDK libraries in combination with each other. Other benefits of these consistent policies include better software quality, productivity, and sustainability.
The xSDK community guidelines (https://xsdk.info/policies) consist of two sets of policies: (1) mandatory policies that affect software quality and ensure compatibility among xSDK libraries and (2) recommended policies that affect additional software quality and sustainability topics. The recommended policies are not necessarily less important than the mandatory policies; instead, recommended policies could be initially categorized as "recommended" to provide existing members with time to achieve compatibility, and these policies could become mandatory at a later time. An xSDK-compatible package is defined as a package that satisfies all of the mandatory policies and can use or be used by at least one other xSDK library.
Allowing for gradual change
To keep the policies up to date and in line with new software developments, the xSDK policy design process allows for policy evolution. The community at large can propose amendments to existing policies as well as suggest new policies. Proposals are then considered and voted on by the community at large and members with existing xSDK packages. The procedure inherently provides a balance between preserving compatibility and gradual obsolescence, as well as between software engineering ideals and what is actually feasible. The latter leads to recommended policies, which---again---are not necessarily less important than mandatory policies but provide some members more time to achieve compatibility before becoming required as policy.
Community of communities
Any software effort will at some point involve interactions with users, and the interactions will lead to a number of known outcomes, including various bugs that range in severity as well as issues with the software features themselves, where they can be too advanced or insufficient for the intended purpose. Similar effects may be observed when the community-defining policies are introduced to a wider audience. Hence, feedback from the participating teams becomes an important aspect for maintaining the relevance of the policies.
For that reason, the xSDK core development team reaches out to other groups to seek additional input and commentary. An example of such outreach is the involvement of application developers and staff members from leadership-class computer facilities. Their feedback sometimes falls outside the prevailing views of the library developers and has a potential for contributing new ideas and out-of-the-box thinking. After discussion, some of this new feedback finds its way into the new community policies' release, leading to an overall improved set of policies.
The xSDK team always welcomes feedback, which can be provided at https://github.com/xsdk-project/xsdk-community-policies.
Author bios
Ulrike Meier Yang leads the Mathematical Algorithms & Computing Group in the Center for Applied Scientific Computing of Lawrence Livermore National Laboratory. She leads the xSDK project in DOE's Exascale Computing Project and the Linear Solvers topical area in the SciDAC FASTMath Institute; she is a developer of the software library hypre. She earned her Ph.D. in computer science from the University of Illinois at Urbana-Champaign. Her research interests are numerical algorithms, parallel computing, and scientific software design.
Piotr Luszczek is a research assistant professor at the Innovative Computing Laboratory in the University of Tennessee. Piotr earned his Ph.D. in computer science from the University of Tennessee, Knoxville. His research interests include benchmarking, numerical linear algebra for high-performance computing, automatic performance tuning, and stochastic performance models. He has over a decade of experience developing HPC numerical software for large-scale, distributed-memory multicore systems with hardware accelerators. Piotr serves as a co-PI on the ECP xSDK project that aims to improve access to world-class software on exascale machines.
