Celebrating the DOE’s Office of Advanced Scientific Computing Research

As per its website, the mission of the Advanced Scientific Computing Research (ASCR) program is “to discover, develop, and deploy computational and networking capabilities to analyze, model, simulate, and predict complex phenomena [that are] important to the Department of Energy (DOE).” ASCR’s prehistory dates back to John von Neumann’s advocacy for increased mathematics and computing activity after the Manhattan Project in the 1940s. Since its inception, ASCR and its predecessor organizations have played a pivotal role in shaping the entire discipline of computational science through investments in basic research, leadership-class facilities and computers, and workforce development programs, among other initiatives. ASCR has also intersected with SIAM in many ways throughout the decades — providing a block grant to support SIAM conferences, funding SIAM members’ research, and much more. The program’s full history is available in ASCR@40: Highlights and Impacts of ASCR’s Programs. To our knowledge, this 115-page report is the most comprehensive history ever written on the topic and is thus highly recommended for scientists, students, and science historians.

The following text is an abridged excerpt from chapter six of ASCR@40, entitled “Lessons Learned and Challenges of the Future.” It has been lightly edited for brevity and provides readers with a thorough overview of previous takeaways, current issues, and potential concerns regarding the future of applied mathematics and computational science. 

Lesson 1: A Compelling and Consistent Vision Can Drive Scientific Revolutions

ASCR and its predecessor organizations have consistently believed that computing is a key driver of science. This sustained commitment drove nearly all of ASCR’s investments, including the development of advanced mathematical techniques, the evolution of computer architecture, the creation of state-of-the-art networking capabilities, an array of innovative computer science concepts, the creation and support of powerful software libraries, and an interdisciplinary workforce. The integration of these capabilities with DOE applications has driven a scientific revolution.

In partnership with scientists at DOE laboratories, ASCR has displayed an admirable focus at key points of technological evolution. It recognized the importance of parallel computing well before the community broadly embraced the technique, and funded critical work to navigate the transition to parallelism. ASCR also appreciated the emerging importance of data science and the critical role of uncertainty quantification as researchers incorporated computer models into decision-making. Collaborations between knowledgeable program managers and the research community have enabled a clear vision and sustained commitment, which are essential to overall progress in these areas.

Lesson 2: Diverse Funding Models Are Required for Diverse and Impactful Outcomes

ASCR has employed a wide variety of different funding models over the years: short- versus long-term, open-ended versus narrowly targeted, large collaborations versus single investigators, and so forth. Graduate and postdoctoral fellowships have helped build the required workforce and attract top scientists to DOE laboratories. Sustained investments in single principal investigators and small teams have enabled fundamental mathematical and computer science advances, while larger, cross-institutional investments have inspired interdisciplinary collaborations.

Software is an essential element of ASCR’s capabilities. Competitive processes drive excellence and innovation, but mechanisms that sustain long-term assets in research and software are also essential. Networking, high-performance computing (HPC) facilities, and HPC platforms require yet another funding model. This broad ecosystem of funding modalities has facilitated ASCR’s greatest successes.

Lesson 3: Workforce Investments Have Been Critical

When working on methodologies that universities had not yet embraced, ASCR invested in workforce development initiatives. The Computational Science Graduate Fellowship has been the most visible and transparently successful effort in this regard, but postdoctoral positions at many laboratories have also been hugely impactful. These investments have staffed DOE laboratories and led to broader achievements in industry and academia. 

Lesson 4: Partnerships Are Essential

Complex challenges require interdisciplinary teams that encompass diverse areas of expertise. These scientific partnerships are best enabled by programmatic partnerships. For example, Scientific Discovery through Advanced Computation (SciDAC) helped overcome organizational barriers and inspired new kinds of science. ASCR has also embraced additional partnerships; for example, the Exascale Computing Project maintains a close partnership with the National Nuclear Security Administration. 

Lesson 5: Testbeds and Platform Access Funding Models Are Important

At points of architectural uncertainty, investing in small testbed systems is critical for understanding the strengths and weaknesses of different designs and making informed decisions about future directions. Larger “early access systems,” wherein users can try full-scale applications and adapt them for the next generation of machines, are particularly beneficial. These steps build confidence in the architectures and allow vendors to learn from early adopters. ASCR has embodied this approach by funding codesign centers on the path to exascale systems.

As growth in the scale of the largest systems outpaces that of widely-available commercial systems, ASCR must continue to invest in the research and development (R&D) that is necessary to build and deploy energy-efficient systems with increased scientific capabilities. To do so, ASCR should keep investing in pathfinding R&D—such as the PathForward program—and non-recurring engineering efforts that are associated with the acquisition of specific systems.

Though scientific computing centers at universities and national laboratories have tried to finance medium- to large-scale systems by charging users for access, this business model does not work for large-scale leadership systems. To advance the state of the art in HPC, the funds to purchase computers must be appropriated and system access should be free for users.

Challenge 1: Technology Disruptions

The vast majority of supercomputer performance improvements in the past several decades have resulted from shrinking microelectronics and faster clock speeds. As these performance drivers come to an end, researchers must find new ways to squeeze additional performance gains from machines. Although the scientific community is embracing simple heterogeneity, future machines will be much more complex. We are entering an era of great change, so strategic clarity and vision—along with sustained investments in creative individuals and high-risk concepts—will be essential. 

Challenge 2: Funding Balance

The rapid emergence of data science and machine learning (ML) in scientific workflows comprises another dramatic shift in the ASCR landscape. With finite dollars, how should ASCR balance support of new areas with continued support of areas of historical strength? Strategic clarity will again be critical when making these hard choices. 

Challenge 3: Software Stewardship

The community has long struggled to identify a good model for sustained support of key elements in the software ecosystem. ASCR must recognize that software is a scientific facility that requires sustained investments in maintenance and support. A simultaneous need for investments in improved software engineering practices reflects ongoing, profound changes in the development and maintenance of modern scientific software. Such technology is often the product of large, dispersed teams and leverages a diverse suite of libraries and tools. These trends complicate development, but researchers can manage them through disciplined software engineering processes like thorough documentation, comprehensive regression suites, and issue tracking.

Challenge 4: Broader Partnerships

The SciDAC program is rightly regarded as a visionary success in building transformative, interdisciplinary partnerships across various scientific domains. There is a growing opportunity for simulation and ML in other areas of the DOE, including organizations that have different value systems and funding models than the Office of Science. How can ASCR partner with this broader group of entities to maximize impact? 

Challenge 5: A Sought-after Workforce

Generational and technological changes will require fresh approaches to workforce issues. Computational scientists have vastly more opportunities in industry and academia than in the past, which increases competition for talent with DOE laboratories.

Challenge 6: New Roles for Computing to Advance Science

For many years, ASCR’s primary goal was to enable more rapid, detailed, and accurate simulations. But in the last several decades, it has broadened its activities to support collaborative technologies and data-centric research and development endeavors. These investments have empowered scientific advances that are quite different from modeling and simulation. As science continues to evolve, ASCR will need to adapt to new roles for scientific computing.

One clear current trend is the explosive growth of ML and artificial intelligence (AI). ASCR researchers are exploring new computing workflows in which simulations and AI work together to generate novel insights. Moving forward, ASCR intends to nurture scientific ML with both fundamental and applied investments. These new workflows will likely drive innovative thinking about the design and usage models for advanced computers.

Another trend is the rapid growth in data streams from DOE user facilities, including light sources, accelerators, and telescopes. The demands associated with the volume and velocity of these data streams require advanced computing. Scientists are increasingly utilizing HPC to analyze experimental data in real time and combine experiments with simulation. As with AI, this trend will presumably motivate fresh thinking about the nature of advanced computing platforms and their integration with data sources and human decision-makers.

Since its establishment, ASCR has played a central role in creating and shaping the field of computational science. This area faces enormous opportunities and challenges in the coming years, and ASCR is poised to continue its pivotal role in advancing the discipline.

SIAM News thanks editor-in-chief Hans Kaper of Georgetown University and Bruce Hendrickson of Lawrence Livermore National Laboratory—who is also co-chair of the Advanced Scientific Computing Advisory Committee Subcommittee on the 40-year History of ASCR—for their review of and contributions to this piece.