Physics in the History of Computing by Peter Freeman (Georgia Institute of Technology)

My focus here is on the organizational history of the U.S. National Science Foundation (NSF) for providing advanced computational resources for the research community. It is a history of push-pull between the advancement in computational capabilities and the need for greater capabilities. It is also a story of interaction specifically between physics and computer science. At NSF this push-pull has led to an on-going tension between the user community (initially physical, geological, and atmospheric sciences and more recently the biological and social sciences) and the research communities on which computer designers rely (primarily physics, electrical engineering, and computer science). That tension has caused NSF’s organizational responsibility for providing advanced computing services to migrate among different directorates and offices from the 1950s to this day. For convenience I’ve divided the story into six time periods, each loosely marked by an important event. I have been a participant in this history from the early 1960’s in several ways, as a primary decision maker about High Performance (HPC) in the early 2000’s, and now as a chronicler of it.

By 1950, when NSF came into existence, the scientific need for HPC was clear to those whose research demanded it. For example, in 1953, the concept of using a computer to perform an experiment in silico was first demonstrated at Los Alamos by Enrico Fermi, John Pasta, and Stanislaw Ulam. One of the first organizational recognitions of the need for computers occurred in the 1955 NSF annual report. In May 1955, the National Science Board decided that NSF should provide computers to universities, although a formal funding program was not established until 1959. A collateral, unplanned effect of some of the resulting grants to computational scientists at least through the 1960’s was to provide support for research which later led to the foundations of computer science. Studies commissioned in 1966 and 1967 led NSF to create the Office of Computing Activities.

By the late 1970s many in the research community felt future scientific advances would be impeded by the lack of advanced computers. In mid-1983, an internal NSF working group, led by Marcel Bardon, Division Director of Physics, and Kent Curtis, Division Director of Computer Research, recommended that NSF provide “supercomputer services for academic research and science education” and support “networks linking universities and laboratories with each other.” Starting in 1985 five awards for NSF supercomputer centers were made to the San Diego Supercomputing Center (SDSC), the National Center for Supercomputing Applications (NCSA) at the University of Illinois at Champaign-Urbana, the John von Neuman Center at Princeton, the Pittsburgh Supercomputer Center, and the Cornell Theory Center. The directors of all the centers were physicists! Except for the Princeton and Cornell centers, all are still in operation today as resources for the scientific community broadly.

In May 1989 Senator Al Gore introduced Senate Bill 1067: “To provide for a coordinated Federal research program to ensure continued United States leadership in high-performance computing.” A later version, known as the High Performance Computing Act of 1991, was enacted on December 9, 1991, and is colloquially known as the “Gore Bill.” It led to the development and funding of the National Research and Education Network (NREN) and advanced HPC. The NREN eventually became the Internet.

The rapid advance of computing, the emergence of the Internet, and the explosion of computer usage created an environment in which NSF often struggled to keep up with the demand and research opportunity in HPC. It seemed as though a new or revised HPC program was barely started before studies and panels were convened to recommend a successor. NSF followed up with two Partnerships for Advanced Computational Infrastructure, called PACIs. Funding was greatly expanded in the period 2000 – 2004. The first awards for terascale computers were made in 2001. By 2004 it was clear to most observers that, for many scientific problems, it was essential to have available the most advanced computational infrastructure possible; it was no longer a discretionary choice in order to be competitive.

Ultimately, NSF created a new Office of Cyberinfrastructure reporting directly to the NSF Director. The Office has been reorganized several times, but to this day NSF continues to provide academic researchers with the latest computational resources. The organizational and programmatic changes at NSF in providing HPC for the general scientific community are a result of two intertwined forces: the continuing, even accelerating, pace of change in the technologies available and the ability of the general scientific community to utilize them. The pace of change is a well-known story. On the other hand, the adoption and utilization of new technology often takes much longer. NSF’s organizational changes, then, are often a direct result of the push-pull between technological advance and scientific need. For the past thirty years the organizational home of HPC has oscillated between the Office of the Director and the Computer and Information Science and Engineering (CISE) Directorate in response to competing demands. As advocacy intensifies for more service for the scientific community in the HPC arena (which they believe the Director will ensure) and an increased focus on utilization of what is available (which leads the Director to move it back to CISE) the office migrates. The broad history of computing and NSF affords a fairly rich context for further historical research.

The above synopsis is based on Chapter 10 of Computing and the National Science Foundation, 1950 – 2016: Building a Foundation for Modern Computing, Peter A. Freeman, W. Richards Adrion, and William Aspray, ACM Books, 2019, New York.

On the Status of Landauer’s Principle by Katherine Robertson (University of Birmingham)

Maxwell’s demon is a creature who cunningly violates the second law of thermodynamics. In what sense is such a demon possible? Whilst thermodynamics legislates against such a creature, the demon looks eminently possible according to the underlying classical or quantum dynamics: Poincare’s recurrence theorem and Loschmidt’s reversibility objection reveal that entropy can decrease in certain situations.

The orthodoxy is that Maxwell’s demon is vanquished by Landauer’s principle, according to which there is an entropy cost to reset the demon’s memory - a vital step in the cyclic process that supposedly leads to a violation of the second law. But the status of Landauer’s principle is controversial: some take it as obviously true, others (such as John Norton) have criticised the proofs of this principle.

In this talk, I clarify the status of Landauer’s principle. First I discuss which assumptions are required to establish Landauer’s principle, and argue that establishing to which theory (thermodynamics, statistical mechanics or quantum mechanics) these principles belong reveals the status of Landauer’s principle. I then consider one of Norton’s counterexamples to Landauer’s principle, and discuss how it depends on certain views about the physical implementation of computation.

Simulation Model Skill in Cosmology by Eric Winsberg (University of South Florida)

What role can/could simulation play in supporting, puzzle-solving, modifying, disconfirming, or falsifying ΛCDM and its competitors? I review some of the problems cosmologists have solved or hope to solve using computer simulation, and examine some of problems and successes that have emerged. I draw some conclusions regarding the kind of simulation model skill we should expect to find in Cosmology. Simulation models have been used for over 50 years now to test and explore the Lambda Cold Dark Matter model of cosmology and its precursors and rivals. The approach has been to simulate the gravitational evolution of CDM in order to predict the evolution of structure in visible matter. But such simulations are highly non-trivial, and give rise to many overlapping Duhem-type problems. But not all such simulations have had the same epistemic goals, and hence not all have been subject to the same problems. My talk explores various approaches to tackling the various problems that arise.

Cosmology in Silico by Marie Gueguen (University of Pittsburgh)

Computer simulations constitute an indispensable tool of contemporary cosmology. They are necessary to extract predictions from cosmological models, to design the observational surveys that will collect the data thanks to which models are assessed, to supplement sparse or non-existing observations. Their ubiquity at every stage of the scientific inquiry must be met with a rigorous methodology for evaluating when simulations faithfully track the physical consequences of the model, especially when expensive observational facilities are built based on simulations-based arguments. Yet, a few astrophysicists (e.g. Melott et al [1], Baushev et al [2], van den Bosch et al [3] have expressed their concerns with respect to current methods for assessing the reliability of simulations, i.e., convergence studies and code comparisons. Consider, for instance, one of the main sources of artefacts in N-body simulations, the fact that dark matter is substituted by fewer, but more massive particles, due to limited resolution. Such an idealization exposes simulations to discreteness-driven effects-e.g., collisional effects, or two-body relaxation. Code comparisons, searching for ‘robust’ properties that remain the same across different codes, do not allow to diagnose such artefacts, based on a common assumption. Convergence studies, on the other hand, look for properties within a given code that resist a change in the value assigned to purely numerical parameters and thus are not sensitive to the specifics of the calibration— e.g., to higher mass or force resolution. Increasing the resolution to diagnose discreteness-driven artefacts does not help however: in the Cold Dark Matter model, structure forms bottom-up, so the first objects to form always contain only a few number of particles. A higher resolution results in a denser environment, from which these first objects condense out with a higher physical density, thereby increasing two-body relaxation effects (Diemand et al [4]). As this example shows, higher resolution is not always better resolution!

Hence, I suggest a method to diagnose artefacts that does not promote the race to ever-increasing resolution but rather facilitate the local evaluation of distinct components of simulations. I refer to this method as that of “crucial simulations”. A ‘crucial simulation’ proposes an idealized, simplified scenario where a physical hypothesis can be tested against a numerical one, by allowing the observation of a prediction drawn from one of the hypotheses and absent from its rival. The observation of the phenomena in the outcome of the simulation then disproves one of the alternatives, thereby confirming the other; i.e., it disproves or confirms the numerical or physical nature of a given property of a simulation outcome. Interesting examples of this method can be found in the literature of the 1980’s, when astrophysicists were developing and testing their P³? codes. Efstathiou and Eastwood [5], for instance, tested their code against two-body relaxation effects by including in their simulations particles of different masses, successfully detecting mass segregation effects. In the simplified scenario upon which crucial simulations are based, moreover, it is more likely to have an analytically tractable solution to compare with the outcomes of simulations, and thus to better understand the impact of specific numerical or physical components. This strategy was used in Efstathiou et al [6]: find a simple case—here, simulating a one-dimensional cloud of particles to verify whether the cloud remains one-dimensional, as it should if the system is collisionless—where exact solutions are available, and test whether these idealized simulations succeed in reproducing the analytically known results. Such a method, given its simplicity, is not computationally expensive and does not have to put in competition with other methods, but rather can be strategically used to complement them when the epistemic opacity of simulations, due to the complexity of the simulated systems and the lack of understanding of how different modules contribute to the final outcome, makes the task of evaluating the reliability of simulations especially difficult.

References

[1] Melott, A. L., Shandarin, S. F., Splinter, R. J., & Suto, Y. (1997). Demonstrating discreteness and

collision error in cosmological N-body simulations of dark matter gravitational clustering. The

Astrophysical Journal Letters, 479(2), L79.

[2] Baushev, A. N., del Valle, L., Campusano, L. E., Escala, A., Muñoz, R. R., & Palma, G. A. (2017).

Cusps in the center of galaxies: a real conflict with observations or a numerical artefact of

cosmological simulations?. Journal of Cosmology and Astroparticle Physics, 2017(05), 042.

[3] Van den Bosch, F. C., & Ogiya, G. (2018). Dark matter substructure in numerical simulations: a tale

of discreteness noise, runaway instabilities, and artificial disruption. Monthly Notices of the Royal

Astronomical Society, 475(3), 4066-4087.

[4] Diemand, J., Moore, B., Stadel, J., & Kazantzidis, S. (2004). Two-body relaxation in cold dark matter

simulations. Monthly Notices of the Royal Astronomical Society, 348(3), 977-986.

[5] Efstathiou, G., & Eastwood, J. W. (1981). On the clustering of particles in an expanding

universe. Monthly Notices of the Royal Astronomical Society, 194(3), 503-525.

[6] Efstathiou, G., Davis, M., White, S. D. M., & Frenk, C. S. (1985). Numerical techniques for large

cosmological N-body simulations. The Astrophysical Journal Supplement Series, 57, 241-260.