By Gary Polhill
Information and Computational Sciences Department, The James Hutton Institute, Aberdeen AB15 8QH, UK.
High-performance computing (HPC) clusters are increasingly being used for agent-based modelling (ABM) studies. There are reasons why HPC provides a significant benefit for ABM work, and to expect a growth in HPC/ABM applications:
- ABMs typically feature stochasticity, which require multiple runs using the same parameter settings and initial conditions to ascertain the scope of the behaviour of the model. The ODD protocol has stipulated the explicit specification of this since it was first conceived (Grimm et al. 2006). Some regard stochasticity as ‘inelegant’ and to be avoided in models, but asynchrony in agents’ actions can avoid artefacts (results being a ‘special case’ rather than a ‘typical case’) and introduces an extra level of complexity affecting the predictability of the system even when all data are known (Polhill et al. 2021).
- ABMs often have high-dimensional parameter spaces, which need to be sampled for sensitivity analyses and, in the case of empirical ABMs, for calibration and validation. The so-called ‘curse of dimensionality’ means that the problem of exploring parameter space grows exponentially with the number of parameters. While ABMs’ parameters may not all be ‘orthogonal’ (i.e. each point in parameter space does not uniquely specify model behaviour – a situation sometimes referred to as ‘equifinality’), diminishing the ‘curse’, the exponential growth means the challenge of parameter search does not need many dimensions before it becomes intractable exhaustively.
- Both the above points are exacerbated in empirical applications of ABMs given Sun et al.’s (2016) observations about the ‘medawar zone’ of model complicatedness in relation to that of theoretical models. In empirical applications, we also may be more interested in knowing that an undesirable outcome cannot occur, or has a very low probability of occurring, requiring more runs with the same conditions. Further, the additional complicatedness of empirical ABM will entail more parameters, and the empirical application will place greater emphasis on searching parameter space for calibrating and validating to data.
HPC clusters are shared computing resources, and it is now commonplace for research organizations and universities to have them. There can be few academic disciplines without some sort of scientific computing requirement – typical applications include particle physics, astronomy, meteorology, materials, chemistry, neuroscience, medicine and genetics. And social science. As a shared resource, an HPC cluster is subject to norms and institutions frequently observed in common-pool resource dilemmas. Users of HPC clusters are asked to request allocations of computing time, memory and long-term storage space to accommodate their needs. The requests are made in advance of the runs being executed; sometimes so far in advance that the calculations form part of the research project proposal. Hence, as a user, if you do not know, or cannot calculate, the resources you will require, you have a dilemma: ask for more than it turns out you really need and risk normative sanctions; or ask for less than it turns out you really need and impair the scientific quality of your research. Normative sanctions are in the job description of the HPC cluster administrator. This can lead to emails such as those in Figure 1.
Figure 1: Example email and accompanying visualization from an HPC cluster administrator reminding users that it is antisocial to request more resources than you will use when submitting jobs.
The ‘managerialist’ turn in academia has been lamented in various articles. Kolsaker (2008), while presenting a nuanced view of the relationship between managerialist and academic modes of working, says that “managerialism represents a distinctive discourse based upon a set of values that justify the assumed right of one group to monitor and control the activities of others.” Steinþórsdóttir et al. (2019) note in the abstract to their article that their results from a case study in Iceland support arguments that managerialism discriminates against women and early-career researchers, in part because of a systemic bias towards natural sciences. Both observations are relevant in this context.
Measurement and control as the tools of managerialist conduct renders Goodhart’s Law (the principle that when a metric becomes a target, the metric is useless) relevant. Goodhart’s Law has been found to have led to bibliometrics now being useless for comparing researchers’ performance – both within and between departments (Fire and Guestrin 2019). We may therefore expect that if an HPC cluster’s administrator has the accurate prediction of computing resource as a target for their own performance assessment, or if they give it as a target for users – e.g. by prioritizing jobs submitted by users on the basis of the accuracy of their predicted resource use, or denying access to those consistently over-estimating requirements – this accuracy will be useless. To give a concrete example, programming languages such as C give the programmer direct control over memory allocation. Hence, were access to an HPC conditional on the accurate prediction of memory allocation requirements, a savvy C programmer would have the (excessive) memory allowance in the batch job submission as a command-line argument to their program, which on execution would immediately request that allocation from the server’s operating system. The rest of the program would use bespoke memory allocation functions that allocated the memory the program actually needed from the memory initially reserved. Similar principles can be used for CPU cycles – if the program runs too quickly, then calculate digits of π until the predicted CPU time has elapsed; and disk space – if too much disk space has been requested, then pad files with random data. These activities waste the programmer’s time, and entail additional use of computing resources with energy cost implications for the cluster administrator.
With respect to the normative statements such as those in Figure 1, Griesemer (2020, p. 77), discussing the use of metrics leading to ‘gaming the system’ in academia generally (the savvy C programmer’s behaviour being an example in the context of HPC usage) claims that “it is … problematic to moralize and shame [such] practices as if it were clear what constitutes ethical … practice in social worlds where Goodhart’s law operates” [emphasis mine]. In computer science, however, there are theoretical (in the mathematical sense of the term) reasons why such norms are problematic over-and-above the social context of measurement-and-control.
The theory of computer science is founded in mathematics and logic, and the work of notable thinkers such as Gödel, Turing, Hilbert, Kolmogorov, Chomsky, Shannon, Tarski, Russell and von Neumann. The growth in areas of computer science (e.g. artificial intelligence, internet-of-things) means that undergraduate degrees have increasingly less space to devote to teaching this theory. Blumenthal (2021, p. 46), comparing computer science curricula in 2014 and 2021, found that the proportion of courses with required modules on computational theory had dropped from 46% to 40%, though the sample size meant this result was not significant (P = 0.09 under a two-population z-test). Similarly, the time dedicated to algorithmics and complexity in CS2013 fell to 28 (of which 19 are ‘tier-1’ – required of every curriculum; and 9 are ‘tier-2’ – in which 80% topic coverage is the stipulated minimum) from 31 in CS2008 (Joint Task Force on Computing Curricula 2013).
One of the most critical theoretical results in computer science is the so-called Halting Problem (Turing 1937), which proves that it is impossible to write a computer program that (in the general case) takes as input another computer program and its input data and gives as output whether the latter program will halt or run forever. The halting problem is ‘tier-1’ in CS2013, and so should be taught to every computer scientist. Rice (1953) generalized Turing’s finding to prove that any ‘non-trivial’ properties of computer programs could not be decided algorithmically. These results mean that the automated job scheduling and resource allocation algorithms in HPC, such as SLURM (Yoo et al. 2003), cannot take a user’s submitted job as input and calculate the computing resources it will need. Any requirement for such prediction is thus pushed to the user. In the general case, this means users of HPC clusters are being asked to solve formally undecidable problems when submitting jobs. Qualified computer scientists should know this – but possibly not all cluster administrators, and certainly not all cluster users, are qualified computer scientists. The power dynamic implied by Kolsaker’s (2008) characterization of a managerialist working culture puts users as a disadvantage, while Steinþórsdóttir et al.’s (2019) observations suggest this practice may be indirectly discriminatory on the basis of age and gender; the latter particularly when social scientists are seeking access to shared HPC facilities.
I emphasized ‘in the general case’ above because in many specific cases, computing resources can be accurately estimated. Sorting a list of strings in alphabetical order, for example is known to grow in execution time with as a function of n log n, where n is the length of the list. Integers can even be sorted in linear time, but with demands on memory that are exponential in the number of bits used to store an integer (Andersson et al. 1998).
However, agent-based modellers should not expect to be so lucky. There are various features that ABMs may implement that make their computing resources difficult (perhaps impossible) to predict:
- Birth and death of agents can render computing time and memory requirements difficult to predict. Indeed, the size of the population and any fluctuation in it may be the purpose of the simulation study. With each agent having memory needed to store its attributes, and execution time for its behaviour, if the maximum population size of a specific run is not predictable from its initial conditions and parameter settings without first running the model, then computing resources cannot be predicted for HPC job submission.
- A more dramatic corollary of birth and death is the question of extinction – i.e. where all agents die before they can reproduce. At this point, a run would typically terminate – far sooner than the computing time budgeted.
- Interactions among agents, where the set of other agents with which one agent interacts is not predetermined, will also typically result in unpredictable computing times, even if the time needed for any one interaction is known. In some cases, agents’ social networks may be formally represented using data structures (‘links’ in NetLogo), and if these connections can be created or destroyed as a result of the model’s dynamics, then the memory requirements will typically be unpredictable.
- Memories of agents, where implemented, are most trivially stored in lists that may have arbitrary length. The algorithms implementing the agents’ behaviours that use their memories will have computing times that are a function of the list length at any one time. These lists may not have a predictable length (e.g. if the agent ‘forgets’ some memories) and hence their behavioural algorithms won’t have predictable execution time.
- Gotts and Polhill (2010) have shown that running a specific model with larger spaces led to qualitatively different results than with smaller spaces. This suggests that smaller (personal) computers (such as desktops and laptops) cannot necessarily be used to accurately estimate execution times and memory requirements prior to submitting larger-scale simulations requiring resources only available on HPC clusters.
Worse, a job will typically comprise several runs in a ‘batch’ covering multiple parameter settings and/or initial conditions. Even if the maximum time and memory requirements of any of the runs in a batch were known, there is no guarantee that all of the other runs will use anything like as much. These matters combine to make agent-based modellers ‘antisocial’ users of HPC clusters where the ‘performance’ of the clusters’ users is measured by their ability to accurately predict resource requirements, or there isn’t an ‘accommodating’ relationship between the administrator and researcher. Further, the social environment in which researchers access these resources put early-career and female researchers at a potential systemic disadvantage
The main purpose of making these points is to lay down the foundations for more equitable access to HPC for social scientists, and provide tentative users of these facilities with the arguments they need to develop constructive working arrangements with cluster administrators for them to run their agent-based models on shared HPC equipment.
This work was supported by the Scottish Government Rural and Environment Science and Analytical Services Division (project reference JHI-C5-1)
Andersson, A., Hagerup, T., Nilsson, S. and Raman, R. (1998) Sorting in linear time? Journal of Computer and System Sciences 57, 74-93. https://doi.org/10.1006/jcss.1998.1580
Blumenthal, R. (2021) Walking the curricular talk: a longitudinal study of computer science departmental course requirements. In Lu, B. and Smallwood, P. (eds.) The Journal of Computing Sciences in Colleges: Papers of the 30th Annual CCSC Rocky Mountain Conference, October 15th-16th, 2021, Utah Valley University (virtual), Orem, UT. Volume 37, Number 2, pp. 40-50.
Fire, M. and Guestrin, C. (2019) Over-optimization of academic publishing metrics: observing Goodhart’s Law in action. GigaScience 8 (6), giz053. https://doi.org/10.1093/gigascience/giz053
Gotts, N. M. and Polhill, J. G. (2010) Size matters: large-scale replications of experiments with FEARLUS. Advances in Complex Systems 13 (4), 453-467. https://doi.org/10.1142/S0219525910002670
Griesemer, J. (2020) Taking Goodhart’s Law meta: gaming, meta-gaming, and hacking academic performance metrics. In Kippmann, A. and Biagioli, M. (eds.) Gaming the Metrics: Misconduct and Manipulation in Academic Research. Cambridge, MA, USA: The MIT Press, pp. 77-87.
Grimm, V., Berger, U., Bastiansen, F., Eliassen, S., Ginot, V., Giske, J., Goss-Custard, J., Grand, T., Heinz, S. K., Huse, G., Huth, A., Jepsen, J. U., Jørgensen, C., Mooij, W. M., Müller, B., Pe’er, G., Piou, C., Railsback, S. F., Robbins, A. M., Robbins, M. M., Rossmanith, E., Rüger, N., Strand, E., Souissi, S., Stillman, R. A., Vabø, R., Visser, U. and DeAngelis, D. L. (2006) A standard protocol for describing individual-based and agent-based models. Ecological Modelling 198, 115-126. https://doi.org/10.1016/j.ecolmodel.2006.04.023
(The) Joint Task Force on Computing Curricula, Association for Computing Machinery (ACM) IEEE Computer Society (2013) Computer Science Curricula 2013: Curriculum Guidelines for Undergraduate Degree Programs in Computer Science. https://doi.org/10.1145/2534860
Kolsaker, A. (2008) Academic professionalism in the managerialist era: a study of English universities. Studies in Higher Education 33 (5), 513-525. https://doi.org/10.1080/03075070802372885
Polhill, J. G., Hare, M., Bauermann, T., Anzola, D., Palmer, E., Salt, D. and Antosz, P. (2021) Using agent-based models for prediction in complex and wicked systems. Journal of Artificial Societies and Social Simulation 24 (3), 2. https://doi.org/10.18564/jasss.4597
Rice, H. G. (1953) Classes of recursively enumerable sets and their decision problems. Transactions of the American Mathematical Society 74, 358-366. https://doi.org/10.1090/S0002-9947-1953-0053041-6
Steinþórsdóttir, F. S., Brorsen Smidt, T., Pétursdóttir, G. M., Einarsdóttir, Þ, and Le Feuvre, N. (2019) New managerialism in the academy: gender bias and precarity. Gender, Work & Organization 26 (2), 124-139. https://doi.org/10.1111/gwao.12286
Sun, Z., Lorscheid, I., Millington, J. D., Lauf, S., Magliocca, N. R., Groeneveld, J., Balbi, S., Nolzen, H., Müller, B., Schulze, J. and Buchmann, C. M. (2016) Simple or complicated agent-based models? A complicated issue. Environmental Modelling & Software 86, 56-67. https://doi.org/10.1016/j.envsoft.2016.09.006
Turing, A. M. (1937) On computable numbers, with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society s2-42 (1), 230-265. https://doi.org/10.1112/plms/s2-42.1.230
Yoo, A. B., Jette, M. A. and Grondona, M. (2003) SLURM: Simple Linux utility for resource management. In Feitelson, D., Rudolph, L. and Schwiegelshohn, U. (eds.) Job Scheduling Strategies for Parallel Processing. 9th International Workshop, JSSPP 2003, Seattle, WA, USA, June 2003, Revised Papers. Lecture Notes in Computer Science 2862, pp. 44-60. Berlin, Germany: Springer. https://doi.org/10.1007/10968987_3
Polhill, G. (2022) Antisocial simulation: using shared high-performance computing clusters to run agent-based models. Review of Artificial Societies and Social Simulation, 14 Dec 2022. https://rofasss.org/2022/12/14/antisoc-sim
© The authors under the Creative Commons’ Attribution-NoDerivs (CC BY-ND) Licence (v4.0)