Pith. sign in

REVIEW 4 major objections 5 minor 26 references

How scarce research compute is allocated is policy, not just engineering—and three gaps decide whether federation helps or harms.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 12:07 UTC pith:KBWOTALA

load-bearing objection Solid UK-facing roadmap: measure occupancy vs utilisation, tune Slurm before replacing it, and don’t open unmetered federation without placement rules—strongest on the first two, sandbox-thin on Braess-style local harm. the 4 major comments →

arxiv 2607.28290 v1 pith:KBWOTALA submitted 2026-07-30 cs.DC

FAIR-Compute: A Roadmap for Fair and Efficient Allocation of Federated Digital Research Infrastructure

Konstantinos (Kostas) E. Zachariadis , Ahmed Sayed , Dimitris Fotakis , Angeliki Mathioudaki , Wan Shuen Siaw This is my paper
classification cs.DC
keywords federated HPCcompute allocationmechanism designSlurm schedulingoccupancy versus utilisationcongestion gamesfairsharedigital research infrastructure
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This report argues that once a shared high-performance computing system must decide who runs, when, and on what evidence, its scheduler settings stop being a purely technical choice and become policy. Drawing on UK and international practice, a mechanism-design model of strategic user reports, and simulations on public workload traces plus synthetic stress tests, it finds three recurring results. Current records track occupancy (resources reserved) rather than utilisation (useful work done), so funders cannot tell whether capacity delivers productive value. Simple, transparent scheduling heuristics, once multi-objective-tuned, come close to a full-information offline benchmark, so the near-term gain is to tune and instrument existing schedulers rather than replace them. Unmanaged federation behaves like selfish road routing: it can cut average delay while concentrating load on receiving systems, so coordinated placement rules are needed. The practical roadmap therefore starts with unified measurement and an occupancy-versus-utilisation KPI, then tuning guidance, portable exchange rates, and governed job mobility.

Core claim

Across landscape evidence, a strategic-reporting model, and simulations, three results recur: allocation data measure occupancy not utilisation, so productive value is unanswerable; tuned transparent heuristics approach an offline full-information welfare benchmark, so tune before replacing; and unmanaged federation minimises average delay like uncoordinated routing while concentrating load—and potential harm—on receiving systems, requiring coordinated placement rules.

What carries the argument

A descriptive Slurm-style priority scheduler paired with an offline full-information welfare benchmark (weighted flow time / value) and a federation-as-congestion-game lens: placement policies from naive lowest-completion routing through switching-cost tolls, minimum-gain thresholds, and utilisation protection, scored on efficiency, delay, and fairness under actual versus requested walltime occupation.

Load-bearing premise

That synthetic high-contention traces built from low-contention public US months, on two abstract systems without real data movement, are enough to ground UK federation policy—including the warning about local harm from selfish placement.

What would settle it

Replay the same scheduling and federation policies on a high-contention, pseudonymised real UK multi-site job log and check whether tuned heuristics still sit near the offline benchmark and whether uncoordinated placement still concentrates load without protecting local users.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A mandatory shared job-telemetry schema is the foundation: without it, efficiency KPIs, ex-post verification, and value-aware pilots cannot be evaluated.
  • Operators should publish multi-objective Slurm weight guidance rather than mandate new schedulers.
  • Portable allocation needs domain-workload exchange rates before open cross-system mobility.
  • Federated job movement should use minimum-gain, utilisation ceilings, or switching tolls instead of pure greedy routing.
  • Occupancy-versus-utilisation reporting would expose over-reservation and support energy and carbon tracking per unit of research output.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If measurement is truly foundational, delaying a shared schema will quietly block every later mechanism the roadmap proposes, including value-aware priority and carbon accounting.
  • The Braess-style analogy implies that adding more free cross-site channels without tolls could worsen local service even as global average wait falls—testable once real multi-site traces exist.
  • Incentive-compatible urgency reporting is the unsolved hinge for any production value-aware pilot; without it, self-declared deadlines will recreate walltime padding in a new form.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. FAIR-Compute frames UK federated DRI allocation as a joint mechanism-design, transport-economics, and scheduling problem. Combining a landscape review (EuroHPC, WLCG, ACCESS, JASMIN, DiRAC), stakeholder interviews, an N=10 demand-side survey, a strategic-reporting model, and simulations on Fresco/Anvil plus synthetic stress tests, it advances three recurring claims: (i) current records measure occupancy not utilisation, so productive value cannot be assessed; (ii) multi-objective-tuned transparent Slurm-like heuristics approach an offline full-information welfare ILP, so near-term gains come from tuning and instrumentation rather than replacement; (iii) unmanaged federation behaves like uncoordinated selfish routing—cutting average delay while concentrating load on receivers—so coordinated placement (tolls, minimum-gain, utilisation protection) is needed before unmetered mobility. Eight FAIRC roadmap recommendations follow, with measurement (FAIRC-1/2) as the foundation.

Significance. If the three recurring results hold under real UK contention, the paper supplies a concrete, sequenced NFCS roadmap that correctly prioritises measurement and low-risk scheduler tuning over speculative scheduler replacement, and that imports a useful congestion-game lens into federated HPC policy. Strengths include explicit triangulation across operator testimony and simulation, honest labelling of synthetic sandbox results and ILP optimality gaps, a public federation simulator (Appendix F), and recommendations written in implementable Must/Should/Could form with FTE/cost estimates. The occupancy-versus-utilisation and “tune before replace” strands are the most transferable contributions; the federation-coordination imperative is the highest-stakes and least externally validated claim.

major comments (4)
  1. [§5.5–5.6, FAIRC-5, Tables 11–13, Appendix D] §5.5–5.6 and FAIRC-5: The load-bearing claim that unmanaged federation requires coordinated placement before unmetered mobility rests on Experiment 4 alone—two abstract 300-node systems, 30-day synthetic traces, a fixed switching-cost stand-in, no inter-site data movement or authentication, and single-window point estimates. The text itself states that local-user harm is not directly quantified and that naive currently looks best on aggregate delay (Table 11–13; Appendix D). Either add a direct local-user slowdown/Jain metric on the receiving system’s home jobs under naive vs coordinated rules, or downgrade FAIRC-5 from a coordination imperative to a pilot contingent on FAIRC-1 UK-trace validation. As written, the Braess analogy is directional evidence, not yet a policy proof.
  2. [§4.3, §5.3, Table 7] §4.3 and Table 7 (Experiment 2): The offline ILP is the welfare benchmark against which “heuristics are good enough” is judged, yet the solver often terminates with a nonzero optimality gap and heuristics sometimes match or beat it on raw wait. The caveat in §4.3 is acknowledged but under-used in the executive claim. Report the optimality gap (or best-bound gap) beside every offline row, and restate the claim as “tuned heuristics are competitive with the best feasible offline solution found,” not as proximity to a proven full-information optimum. Without that, the “tune don’t replace” headline over-reaches the computation.
  3. [§5.1, §5.6, Tables 5–8] §5.1 and §5.6: All contention results that drive Experiments 1–3 and the federation stress tests use synthetic bursty-fill traces derived from low-contention Anvil months (real-trace utilisation pinned ≈0.25 in Table 8). Volume multipliers and fill injection are free parameters. Sensitivity of the Pareto weight sets (Table 5–6), the heuristic–ILP gap (Table 7), and transfer volumes (Table 16) to the fill multiplier and to alternative cluster counts should be reported; otherwise the policy levers and “close to offline” conclusions are tied to one sandbox calibration. This is the external-validity hinge the authors already flag as motivating FAIRC-1—make the dependence quantitative in the main text.
  4. [§4.2, §5.4, Appendix A, FAIRC-8/3] §4.2 and Experiment 3: The strategic-reporting utility (Eq. 2) and penalty mechanisms (Appendix A) are presented as the analytical spine for FAIRC-8 and FAIRC-3, but “the strategic-reporting mechanism itself is not exercised in the present simulations.” Experiment 3 shows runtime externalities and value loss under value-blind rules, which is related but not a test of incentive-compatible walltime or urgency reporting. Either run a minimal strategic-padding experiment (users best-responding to kill-at-walltime / fairshare penalties) or clearly separate the mechanism-design programme as future work so that FAIRC-8 is not read as simulation-backed when it is practice-codification plus untested theory.
minor comments (5)
  1. [Executive Summary, §2.2] N=10 survey (§2.2) is correctly labelled exploratory, but the Executive Summary’s “three strands of evidence” phrasing can be read as equal weight. Soften survey citations in the Exec. Summary to “corroborative operator-facing themes.”
  2. [Figures 5, 7; Tables 9–13] Figure 5 and Figure 7 need axis units and a one-line reading guide in the caption; several tables report point estimates without stating the job-window size in the table note (only in §5.6).
  3. [Appendix C] QoS-tier proxy cut-points (Appendix C, Table 14) are ad hoc; a one-sentence robustness note that conclusions are insensitive to cut-points would help, as claimed in the appendix but not shown.
  4. [§6] Recommendation numbering jumps FAIRC-1,2,4,7,6,8,3,5 in implementation order—fine for a roadmap, but add a one-line dependency diagram or explicit “depends on” column so readers do not assume numeric order is priority order.
  5. [§1, throughout] Minor typos/consistency: “becomespolicy” spacing in the Intro; “use-it-or-lose-it” hyphenation varies; arXiv date and “Submitted July 2026” are fine but ensure reference access dates are uniform.

Circularity Check

0 steps flagged

No significant circularity: empirical/simulational comparisons against independent benchmarks, not definitions restated as predictions.

full rationale

FAIR-Compute’s load-bearing claims are landscape triangulation plus controlled simulation comparisons, not a first-principles derivation that collapses into its inputs. The offline welfare benchmark (weighted flow-time/value ILP, §4.3 Eq. 3) is an independent full-information reference; FIFO/Slurm-like/tuned heuristics are scored against it rather than fitted to reproduce it (§5.3). Multi-objective weight search (NSGA-II) maps trade-offs among utilisation, wait, slowdown, and fairness; it does not claim to predict a quantity already used as the fit target (§5.2). QoS tiers are explicitly labelled inferred ACCESS-style proxies from annualised usage, with sensitivity disclaimed (Appendix C). Federation results compare placement policies under synthetic contention and invoke Braess’s paradox only as an interpretive congestion-game analogy, not as a uniqueness theorem or self-cited forced result (§5.5–5.6). Strategic-reporting utility (Eq. 2, Appendix A) is mechanism-design framing and is not exercised as a circular ‘prediction’ in the simulations. External citations (Slurm, ACCESS, WLCG/HEPScore, pymoo, PyJobShop) are operational or methodological scaffolding, not author-overlapping uniqueness imports that forbid alternatives. Weak external validity of the synthetic sandbox is a correctness/generalisation issue, not circularity by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central policy claims rest on standard scheduling and congestion-game machinery, operator testimony, and sandbox simulations—not on new physical entities. Load-bearing modelling choices are the descriptive Slurm-like priority score, linear waiting disutility, synthetic contention generation from Anvil structure, inferred ACCESS-style QoS proxies, a stylised two-site federation with switching tolls, and a synthetic scientific-value/deadline model used only in Experiment 3. Free parameters are administrative weights, toll/threshold settings, and clustering/generation choices rather than fitted universal constants.

free parameters (6)
  • Slurm-like priority weights (age, fairshare, size, QoS, ...) = Pareto set on synthetic filled workload (Table 5); single-objective runs in Table 6
    Treated as searchable policy levers via NSGA-II; Pareto policies (e.g. P1/P5/P6/P8) are outcomes of finite evolutionary search on chosen metrics, not unique optima.
  • Federation switching-cost toll = fixed + 10% of runtime (Appendix D)
    Fixed component plus 10% of selected runtime on remote execution; stands in for network/transfer cost without modelling data movement.
  • Minimum-gain threshold θ and utilisation ceiling U_max
    Coordination dials in Experiment 4 that filter transfers; specific numeric settings drive transfer counts and wait trade-offs.
  • ACCESS-style QoS-tier cut-points from annualised usage = ≤4e5; (4e5,1.5e6]; (1.5e6,3e6]; >3e6 core-hour proxy (Table 14)
    Inferred proxy tiers (Explore/Discover/Accelerate/Maximize) mapped from annualised core-hour demand because true project type is absent in Fresco.
  • Synthetic workload volume multipliers and bursty-fill injection
    Contention is created by regenerating jobs from learned clusters and adding volume; results on policy bite depend on this controlled pressure.
  • Linear waiting penalty κ_j and job values V_j in welfare/utility model = Experiment 3: 2388 value units over 500 jobs
    Offline benchmark and user utility use V_j − κ_j W_j; deadline-value experiment uses a synthetic full-deadline value metric (2388 units / 500 jobs).
axioms (6)
  • domain assumption Production schedulers rank jobs by a weighted linear priority score over age, fairshare, QoS, size, partition (Eq. 1), with backfill of smaller jobs that do not delay the head job.
    Descriptive model of Slurm-like practice used throughout simulations (§4.1).
  • domain assumption Users face uncertain runtimes and choose reported walltime as self-selected insurance trading kill/preempt risk against queue delay and ex-post penalties (Eq. 2).
    Mechanism-design framing for padding; largely not exercised in the reported simulations (§4.2, Appendix A).
  • domain assumption Offline full-information scheduling maximising sum (V_j − κ_j W_j) is a meaningful welfare benchmark for practical rules even when the solver returns a small optimality gap.
    Experiment 2 comparison basis; paper notes solutions may not be certified global optima (§4.3).
  • ad hoc to paper Uncoordinated cross-system placement is analogous to unpriced selfish routing in congestion games, so Braess-like concentration risks motivate coordinated tolls/thresholds.
    Interpretive lens for Experiment 4; paper admits local harm is not directly quantified and naive looks best on aggregate delay (§5.5).
  • ad hoc to paper Synthetic traces that preserve Anvil job-structure clusters but amplify contention are valid instruments for evaluating allocation policy where real months are low-contention.
    Explicit methodological stance because UK logs were unavailable (§§2.4, 5.1, 5.6).
  • domain assumption Occupancy-based accounting systematically diverges from productive utilisation and is the right efficiency gap for national KPIs.
    Recurring operator theme elevated to FAIRC-2; supported qualitatively, not by a single causal estimate (§3.2–3.3).
invented entities (3)
  • FAIRC-1..8 recommendation bundle (unified schema, occ/util KPI, PVR pilot, etc.) no independent evidence
    purpose: Package evidence into an NFCS roadmap with Must/Should/Could priorities and FTE/cost estimates.
    Policy artifacts derived from the study, not physical entities; PVR is proposed as a future pilot measure of non-increasing job value over time.
  • Projected Value Remaining (PVR) prioritisation measure no independent evidence
    purpose: Pilot value-aware scheduling that accounts for scientific value lost when time-sensitive jobs wait (FAIRC-3).
    Introduced as a could-have pilot; Experiment 3 uses a synthetic deadline-value model as motivation, not a validated operational PVR implementation.
  • Inferred ACCESS-style QoS-tier proxy per account no independent evidence
    purpose: Supply account-scale heterogeneity in the simulator when true project types are missing from Fresco.
    Engineering construct for simulation realism (Appendix C), labelled as proxy not ground truth.

pith-pipeline@v1.2.0-daily-grok45 · 27469 in / 4613 out tokens · 81713 ms · 2026-07-31T12:07:19.947752+00:00 · methodology

0 comments
read the original abstract

As demand for high-performance computing (HPC), high-throughput computing and data storage grows, the way scarce compute is allocated -- not just how much exists -- has become a decisive factor in the productivity of UK research. FAIR-Compute studies allocation in a federated Digital Research Infrastructure (DRI) as a problem at the intersection of algorithmic game theory, transport economics and HPC scheduling. Our central observation is simple: once a shared system must decide who runs, when and under what evidence, its scheduler settings cease to be a purely technical matter and become policy. The project combined three strands of evidence: a landscape review of UK and international allocation practice (including EuroHPC, WLCG, ACCESS, JASMIN and DiRAC) supported by a stakeholder survey; a mechanism-design model of allocation under strategic and uncertain user reports; and a simulation study built on the public Fresco/Anvil workload trace and controlled synthetic stress tests. Three results recur across all three strands. First, allocation records measure occupancy (resources reserved) rather than utilisation (useful work done), so the system cannot currently answer the question UKRI and DSIT most want answered -- whether resources deliver productive value. Second, simple, transparent scheduling heuristics perform close to an offline full-information benchmark once tuned, so the near-term opportunity is to tune and instrument existing schedulers rather than replace them. Third, federation is genuinely valuable but behaves like uncoordinated routing when left unmanaged: it can minimise average delay while quietly concentrating load -- and potentially harm -- on the receiving system. This mirrors a classic effect from road-traffic economics (Braess's paradox), where letting everyone independently pick the fastest route can leave the whole network worse off.

Figures

Figures reproduced from arXiv: 2607.28290 by Ahmed Sayed, Angeliki Mathioudaki, Dimitris Fotakis, Konstantinos (Kostas) E. Zachariadis, Wan Shuen Siaw.

Figure 1
Figure 1. Figure 1: Descriptive statistics of the Anvil workload: bursty temporal demand (left) and a [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Clustering diagnostics used to choose and validate the number of workload groups [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: SHAP feature importance for the surrogate workload classifier. The learned clusters [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Account-level fairness (Jain index) across scheduling rules. Under low contention (left) [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Job-level waiting times under runtime perturbation. Red points are jobs whose [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Achieved scientific value by scheduler. The offline benchmark preserves the full [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-system transfers over time under naive versus switching-cost routing (asym [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

26 extracted references · 1 linked inside Pith

  1. [1]

    ACCESS Project Types.https://allocations.access-ci.org/ project-types, 2026

    ACCESS. ACCESS Project Types.https://allocations.access-ci.org/ project-types, 2026. Accessed: 2026-07-01

    2026

  2. [2]

    ACCESS Resources.https://allocations.access-ci.org/resources, 2026

    ACCESS. ACCESS Resources.https://allocations.access-ci.org/resources, 2026. Accessed: 2026-07-01

    2026

  3. [3]

    Pymoo: Multi-objective optimization in python.Ieee access, 8:89497–89509, 2020

    Julian Blank and Kalyanmoy Deb. Pymoo: Multi-objective optimization in python.Ieee access, 8:89497–89509, 2020

    2020

  4. [4]

    White paper on China’s computing power development index

    China Academy of Information and Communications Technology (CAICT). White paper on China’s computing power development index. English translation (2024), Center for Security and Emerging Technology, 2022.https://cset.georgetown.edu/wp-content/ uploads/t0581_china_compute_index_2022_EN.pdf

    2024

  5. [5]

    A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE transactions on evolutionary computation, 6(2):182–197, 2002

    Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii.IEEE transactions on evolutionary computation, 6(2):182–197, 2002

    2002

  6. [6]

    The impact of epsrc’s investments in high performance comput- ing infrastructure: Final report

    EPSRC. The impact of epsrc’s investments in high performance comput- ing infrastructure: Final report. Technical report, UK Research and In- novation, 2019.https://www.ukri.org/wp-content/uploads/2022/07/ EPSRC-050722-ImpactEPSRCInvestmentsHighPerformanceComputingInfrastructure. pdf

    2019

  7. [7]

    Federation of computing infrastructures

    ETP4HPC. Federation of computing infrastructures. Technical report, European Technol- ogy Platform for High-Performance Computing, 2025

    2025

  8. [8]

    Eurohpc federation platform.https://www.eurohpc-ju

    EuroHPC Joint Undertaking. Eurohpc federation platform.https://www.eurohpc-ju. europa.eu/supercomputers/eurohpc-federation-platform_en, 2026. Accessed: 2026- 06-30

    2026

  9. [9]

    FRESCO: Open repository and analysis of system usage data.https: //www.frescodata.xyz/, 2026

    FRESCO Project. FRESCO: Open repository and analysis of system usage data.https: //www.frescodata.xyz/, 2026. Accessed: 2026-05-04

    2026

  10. [10]

    The economic impact of NVIDIA cambridge-1, 2021

    Frontier Economics. The economic impact of NVIDIA cambridge-1, 2021. Estimated economic value approx. £600M (approx. $831M) over ten years

    2021

  11. [11]

    HEPScore: A new CPU benchmark for the WLCG

    Domenico Giordano et al. HEPScore: A new CPU benchmark for the WLCG. InEPJ Web of Conferences (CHEP), 2023. Adopted by WLCG April 2023, replacing HEP-SPEC06; reproducible to better than 1% across x86 and ARM.https://arxiv.org/abs/2306. 08118. 29 FAIR-Compute NFCS Flexible Fund — Final Report

    2023

  12. [12]

    Overview of HPCI.https://www

    High Performance Computing Infrastructure (HPCI). Overview of HPCI.https://www. hpci-office.jp/en/about_hpci/what_is_hpci, 2024. Accessed: 2026-07-17

    2024

  13. [13]

    Open ondemand: A web-based client portal for hpc centers

    David Hudak, Doug Johnson, Alan Chalker, Jeremy Nicklas, Eric Franz, Trey Dockendorf, and Basil L McMichael. Open ondemand: A web-based client portal for hpc centers. Journal of Open Source Software, 3(25):622, 2018

    2018

  14. [14]

    Pyjobshop: Solving scheduling problems with constraint programming in python.arXiv preprint arXiv:2502.13483, 2025

    Leon Lan and Joost Berkhout. Pyjobshop: Solving scheduling problems with constraint programming in python.arXiv preprint arXiv:2502.13483, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    SHAP: Shapley additive explanations documen- tation.https://shap.readthedocs.io/en/latest/, 2026

    Scott Lundberg and SHAP contributors. SHAP: Shapley additive explanations documen- tation.https://shap.readthedocs.io/en/latest/, 2026. Accessed: 2026-05-12

    2026

  16. [16]

    Fresco: A public multi-institutional dataset for understanding hpc system behavior and depend- ability

    Joshua McKerracher, Preeti Mukherjee, Rajesh Kalyanam, and Saurabh Bagchi. Fresco: A public multi-institutional dataset for understanding hpc system behavior and depend- ability. InPractice and Experience in Advanced Research Computing 2025: The Power of Collaboration, pages 1–6. 2025

    2025

  17. [17]

    PyJobShop Documentation.https://pyjobshop.org/stable/ index.html, 2026

    PyJobShop Developers. PyJobShop Documentation.https://pyjobshop.org/stable/ index.html, 2026. Accessed: 2026-07-01

    2026

  18. [18]

    pymoo: Multi-objective Optimization in Python.https://pymoo

    pymoo Developers. pymoo: Multi-objective Optimization in Python.https://pymoo. org/, 2026. Accessed: 2026-07-01

    2026

  19. [19]

    High performance computing development in China: A brief review and perspectives.Computing in Science & Engineering, 21(1):6–16, 2018

    Depei Qian and Zhongzhi Luan. High performance computing development in China: A brief review and perspectives.Computing in Science & Engineering, 21(1):6–16, 2018. doi:10.1109/MCSE.2018.2875367

    arXiv 2018

  20. [20]

    Slurm workload manager documentation, 2026

    SchedMD. Slurm workload manager documentation, 2026. URLhttps://slurm.schedmd. com/documentation.html. Accessed: 2026-03-14

    2026

  21. [21]

    Anvil-system architec- ture and experiences from deployment and early user operations

    X Carol Song, Preston Smith, Rajesh Kalyanam, Xiao Zhu, Eric Adams, Kevin Colby, Patrick Finnegan, Erik Gough, Elizabett Hillery, Rick Irvine, et al. Anvil-system architec- ture and experiences from deployment and early user operations. InPractice and experience in advanced research computing 2022: Revolutionary: Computing, connections, you, pages 1–9. 2022

    2022

  22. [22]

    Met office supercomputing 2020+ programme: Accounting officer assess- ment

    UK Government. Met office supercomputing 2020+ programme: Accounting officer assess- ment. Technical report, Government Major Projects Portfolio, GOV.UK, 2022. Net present social value £13.74bn (25% optimism bias); 9:1 cost–benefit ratio against the do-nothing option

    2020

  23. [23]

    National compute ecosystem town hall

    UK Research and Innovation. National compute ecosystem town hall. UKRI Digital Re- search Infrastructure programme, 2026. June 2026

    2026

  24. [24]

    Waldur: Cloud and hpc management platform.https://waldur.com/, 2026

    Waldur. Waldur: Cloud and hpc management platform.https://waldur.com/, 2026. Accessed: 2026-06-30

    2026

  25. [25]

    Memorandum of understanding for the worldwide lhc comput- ing grid.https://wlcg.web.cern.ch/organisation-mou/sample-mou, 2020

    WLCG Collaboration. Memorandum of understanding for the worldwide lhc comput- ing grid.https://wlcg.web.cern.ch/organisation-mou/sample-mou, 2020. Accessed: 2026-06-30

    2020

  26. [26]

    Toward dynamically controlling slurm’s classic fairshare algorithm

    Jason Yalim. Toward dynamically controlling slurm’s classic fairshare algorithm. InPrac- tice and Experience in Advanced Research Computing 2020: Catch the Wave, pages 538–

    2020