pith. sign in

arxiv: 2604.26576 · v1 · submitted 2026-04-29 · 💻 cs.DC

MPI Malleability Validation under Replayed Real-World HPC Conditions

S. Iserte , M. Madon , G. Da , J. Pierson , A. J. Pe\~na This is my paper

Pith reviewed 2026-05-07 10:56 UTC · model grok-4.3

classification 💻 cs.DC
keywords MPI malleabilitydynamic resource managementworkload replayHPC clustersparallel efficiencyresource utilizationjob scheduling
0
0 comments X

The pith

Replaying real HPC workload logs on a production supercomputer shows efficiency-aware MPI malleability shortens malleable job times by 27% without delaying baseline workloads.

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

The paper develops a method to validate dynamic resource management techniques by replaying actual workload logs on real HPC hardware, adapting the logs to match the target cluster's configuration. This directly tackles administrator skepticism that malleability and similar approaches only work in simulations. Testing occurred on a 125-node partition where malleable applications could resize their MPI process counts. Results indicate that guiding malleability by parallel efficiency cut the overall time for the malleable portion of the workload by 27 percent, preserved resource utilization, and left the baseline non-malleable workload's completion time unchanged, even though some individual jobs waited longer in the queue.

Core claim

The authors introduce a replay-based validation method that reproduces real cluster conditions by adapting workload logs to the target HPC infrastructure. When applied to MPI malleability on a 125-node partition, parallel efficiency-aware malleability reduced the malleable workload's execution time by 27% without increasing the completion time of the baseline workload, while maintaining the resource utilization rate despite added queueing delays for certain jobs.

What carries the argument

The workload log replay methodology, which adapts historical job and user data to current cluster conditions to enable realistic validation of malleable MPI applications that dynamically adjust their process count based on observed parallel efficiency.

If this is right

  • Malleable workloads finish earlier when resizing follows parallel efficiency.
  • Non-malleable baseline workloads incur no extra delay from the presence of malleable jobs.
  • Cluster-wide resource utilization remains at the same level as without malleability.
  • Some jobs experience longer queue waits as a side effect of dynamic process adjustment.

Where Pith is reading between the lines

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

  • The replay technique could be reused to test other dynamic resource management methods such as job migration or power capping.
  • Production clusters might first run limited log-replay pilots before enabling malleability cluster-wide.
  • Comparing replay outcomes against pure simulation results would quantify how much additional realism the log replay supplies.
  • Similar efficiency gains could appear on other systems whose workload traces show comparable job-size distributions.

Load-bearing premise

Replaying the workload logs faithfully reproduces the actual job arrival patterns, user behaviors, and cluster conditions on the target system.

What would settle it

Running the identical malleability policy on the same replayed logs but measuring no reduction near 27% in malleable workload time or an increase in baseline workload completion time would falsify the reported performance benefit.

Figures

Figures reproduced from arXiv: 2604.26576 by A. J. Pe\~na, G. Da, J. Pierson, M. Madon, S. Iserte.

Figure 1
Figure 1. Figure 1: DMRlib Application–MPI–Slurm communication. 3.3. Dynamic Resource Management This research enables DRM by leveraging moldability and malleability thanks to the Dynamic Management of Resources Library (DMRlib) [34]. DMRlib is a high-level API that facilitates the adoption of malleability in HPC codes. DMRlib implements a communication layer between the parallel distributed runtime (PDR) and the RMS, driving… view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of job submissions and platform capacity over the days of July 2017 (x-axis). Each color represents a different user. The horizontal line in the bottom graph is the maximum capacity of the platform (𝑀 = 124 × 24 = 2, 976 node-hours per day). 10 0 10 1 10 2 10 3 10 4 Job Execution Time (Minutes) - Log Scale 0.0 0.2 0.4 0.6 0.8 1.0 Cumulative Proportion of Jobs count 1h 10h 1day 2day 3day view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of job execution time in the baseline workload. 4.2. Traditional Users: The Baseline Workload The baseline workload is adapted from the most recent available workload on the Parallel Workload Archive KIT￾FH2-20166 log, recorded from the ForHLR II system located at the Karlsruhe Institute of Technology in Germany. Notice 5 https://www.bsc.es/supportkc/docs/MareNostrum5/overview 6 https://www.cs… view at source ↗
Figure 4
Figure 4. Figure 4: Job sizes in the baseline workload. The distribution is shown by the number of jobs (top) and number of node-hours (bottom), and the cumulative distribution is represented in grey. visualization queue with 21 nodes comprising CPUs and GPUs. We focus on the first queue and exclude the GPU￾enabled. 2. The record is 1.5 years long, and we only have two days of computation available in our testbed. Consequentl… view at source ↗
Figure 5
Figure 5. Figure 5: MPDATA scalability in MN5 view at source ↗
Figure 6
Figure 6. Figure 6: MPDATA reconfiguration times in MN5. are available resources in the cluster (line 10), the policy checks the current parallel efficiency. If the value exceeds a determined threshold (line 11), indicating that the execution may still leverage additional resources, the malleable job will be expanded (line 12). This policy aims to obtain as many resources as possible if no other job in the queue may use them.… view at source ↗
Figure 7
Figure 7. Figure 7: Resource allocation (Y-axis) in each second of the execution (X-axis) for Baseline experiment. Deadline (a) StaticN32 Deadline (b) StaticN16 view at source ↗
Figure 8
Figure 8. Figure 8: Resource allocation (Y-axis) over execution time (X-axis) for the static experiments of the student. the deadline. AlwaysGrow, leverages malleability to reduce the student makespan, but it still needs 36.62 actual hours, which is still beyond the deadline (see Figure 9a). Notably, the “smartest” dynamic configuration, ParEfficiency, needs ≈ 73% the time to complete the student jobs compared to StaticN32 (3… view at source ↗
Figure 9
Figure 9. Figure 9: Resource allocation (Y-axis) over execution time (X-axis) for the dynamic experiments of the student. time, which are the two factors that explain the makespan reduction of the PhD student workload. 5.2. Resource Allocation Rate view at source ↗
Figure 10
Figure 10. Figure 10: Accumulated waiting time throughout the workloads executions. It is patent that the original workload suffers longer de￾lays on average when the PhD student jobs are submitted: the average waiting time increases from 1,725.04 s. to 3,007.24 s. or 4,365.89 s. in the best and worst cases, respectively. This increment of up to ≈ 2.53𝑥 is expected since the system is more saturated with the extra jobs. Beside… view at source ↗
Figure 11
Figure 11. Figure 11: Difference in job waiting times compared to the baseline experiment view at source ↗
Figure 12
Figure 12. Figure 12: Average student’s job completion (waiting + execution) time (Y-axis) for the four experiments (X-axis). the PhD student workload time by 10% compared to Al￾waysGrow. The impact of these policies is also reflected in node￾hour consumption view at source ↗
Figure 13
Figure 13. Figure 13: Individual student job’s Node-hours (Y-axis) grouped by colors for the four experiments (X-axis). Dashed lines represent the average per experiment (Y-axis). S. Iserte et al.: Preprint submitted to Elsevier Page 14 of 22 view at source ↗
read the original abstract

Dynamic Resource Management (DRM) techniques can be leveraged to maximize throughput and resource utilization in computational clusters. Although DRM has been extensively studied through analytical workloads and simulations, skepticism persists among end administrators and users regarding their feasibility under real-world conditions. To address this problem, we propose a novel methodology for validating DRM techniques, such as malleability, in realistic scenarios that reproduce actual cluster conditions of jobs and users by replaying workload logs on a High-performance Computing (HPC) infrastructure. Our methodology is capable of adapting the workload to the target cluster. We evaluate our methodology in a malleability-enabled 125-node partition of the Marenostrum 5 supercomputer. Our results validate the proposed method and assess the benefits of MPI malleability on a novel use case of a pioneer user of malleability (our "PhD Student"): parallel efficiency-aware malleability reduced a malleable workload time by 27% without delaying the baseline workload, although introducing queueing delays for individual jobs, but maintaining the resource utilization rate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes a methodology for validating dynamic resource management techniques such as MPI malleability by replaying real-world workload logs on actual HPC hardware, with workload adaptation to the target cluster. Evaluated on a 125-node malleability-enabled partition of Marenostrum 5, it claims that parallel efficiency-aware malleability reduces malleable workload completion time by 27% without delaying the baseline workload, while introducing queueing delays for some jobs but preserving overall resource utilization.

Significance. If the replay methodology accurately captures real conditions, this provides empirical evidence from production hardware that could help overcome administrator skepticism toward DRM techniques. The use of replayed logs on real infrastructure rather than pure simulation is a strength for credibility and reproducibility.

major comments (2)
  1. [Evaluation methodology] The central 27% reduction claim (abstract) depends on static log replay with fixed submission times reproducing actual cluster behavior. However, this does not model dynamic user responses where shortened job runtimes could prompt earlier follow-on submissions, potentially changing observed time savings, queueing delays, and utilization.
  2. [Abstract and evaluation] Support for the reported outcomes on Marenostrum 5 is limited by insufficient details on experimental setup, controls, statistical significance, and how the workload was adapted to the 125-node partition, as these are load-bearing for validating the malleability benefits.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and indicate the revisions we will make to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Evaluation methodology] The central 27% reduction claim (abstract) depends on static log replay with fixed submission times reproducing actual cluster behavior. However, this does not model dynamic user responses where shortened job runtimes could prompt earlier follow-on submissions, potentially changing observed time savings, queueing delays, and utilization.

    Authors: Our methodology deliberately replays fixed submission times from production logs to reproduce observed cluster conditions on real hardware without introducing unverified assumptions about user behavior. Modeling dynamic responses (e.g., earlier follow-on submissions) would require additional user studies or behavioral models outside the scope of this hardware-validation-focused work. We will add a brief discussion of this limitation and its implications in the revised manuscript. revision: partial

  2. Referee: [Abstract and evaluation] Support for the reported outcomes on Marenostrum 5 is limited by insufficient details on experimental setup, controls, statistical significance, and how the workload was adapted to the 125-node partition, as these are load-bearing for validating the malleability benefits.

    Authors: We agree that additional details are required. In the revised manuscript we will expand the experimental setup section with explicit information on controls, the statistical methods used to assess significance of the reported 27% reduction, and the precise adaptation steps applied to the workload logs for the 125-node partition. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical log-replay validation on real hardware

full rationale

The paper describes an empirical methodology that replays existing workload logs on a 125-node malleability-enabled partition of Marenostrum 5 to measure the effects of MPI malleability. The reported 27% reduction in malleable workload time is an observed outcome of the hardware experiment rather than a quantity derived from equations or fitted parameters. No mathematical derivation chain exists that reduces a claimed prediction back to its own inputs by construction, nor are there load-bearing self-citations, uniqueness theorems, or ansatzes that close on themselves. The evaluation is therefore self-contained against external benchmarks (real hardware runs) and receives a score of 0.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the replay methodology faithfully captures real conditions, with no free parameters or new entities introduced.

axioms (1)
  • domain assumption Replaying workload logs on the target cluster accurately reproduces real-world conditions of jobs and users.
    This is central to claiming the validation is realistic rather than simulated.

pith-pipeline@v0.9.0 · 5491 in / 1102 out tokens · 79964 ms · 2026-05-07T10:56:19.099666+00:00 · methodology

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

41 extracted references · 32 canonical work pages

  1. [1]

    Martínez, J

    H. Martínez, J. Tárraga, I. Medina, S. Barrachina, M. Castillo, J. Dopazo, E. S. Quintana-Ortí, A dynamic pipeline for RNA se- quencing on multicore processors, in: Proceedings of the 20th Eu- ropean MPI Users’ Group Meeting, EuroMPI ’13, Association for Computing Machinery, New York, NY, USA, 2013, pp. 235–240. doi:10.1145/2488551.2488581

    work page doi:10.1145/2488551.2488581 2013

  2. [2]

    doi:10.1007/s42514-024-00203-0

    H.Zhong,X.Pan,Z.He,H.Wang,D.Huang,Z.Chen,GPUacceler- ationforDNAsequencealignmentalgorithmanditsapplication,CCF TransactionsonHighPerformanceComputing7(2)(2025)169–177. doi:10.1007/s42514-024-00203-0

    work page doi:10.1007/s42514-024-00203-0 2025

  3. [3]

    B. Hess, C. Kutzner, D. van der Spoel, E. Lindahl, GROMACS 4: Algorithms for Highly Efficient, Load-Balanced, and Scalable MolecularSimulation,JournalofChemicalTheoryandComputation 4 (3) (2008) 435–447.doi:10.1021/ct700301q

    work page doi:10.1021/ct700301q 2008

  4. [4]

    Vázquez, G

    M. Vázquez, G. Houzeaux, S. Koric, A. Artigues, J. Aguado-Sierra, R. Arís, D. Mira, H. Calmet, F. Cucchietti, H. Owen, A. Taha, E.D.Burness,J.M.Cela,M.Valero,Alya:Multiphysicsengineering simulation toward exascale, Journal of Computational Science 14 (2016) 15–27.doi:10.1016/j.jocs.2015.12.007

    work page doi:10.1016/j.jocs.2015.12.007 2016

  5. [5]

    SERGHEI (SERGHEI-SWE) v1.0: a performance-portable high-performance parallel-computing shallow-water solver for hydrology and environmental hydraulics

    D. Caviedes-Voullième, M. Morales-Hernández, M. R. Norman, I. Özgen Xian, SERGHEI (SERGHEI-SWE) v1.0: a performance- portable high-performance parallel-computing shallow-water solver for hydrology and environmental hydraulics, Geoscientific Model Development16(3)(2023)977–1008.doi:10.5194/gmd-16-977-2023

    work page doi:10.5194/gmd-16-977-2023 2023

  6. [6]

    R.Martínez-Cuenca,J.Luis-Gómez,S.Iserte,S.Chiva,OntheUseof DeepLearningandComputationalFluidDynamicsfortheEstimation of Uniform Momentum Source Components of Propellers, iScience 26 (2023) 1–14.doi:10.1016/j.isci.2023.108297

    work page doi:10.1016/j.isci.2023.108297 2023

  7. [7]

    P.Rosciszewski,A.Krzywaniak,S.Iserte,K.Rojek,P.Gepner,Opti- mizing Throughput of Seq2Seq Model Training on the IPU Platform for AI-accelerated CFD Simulations, Future Generation Computer Systems 143 (2023) 149–162.doi:10.1016/j.future.2023.05.004

    work page doi:10.1016/j.future.2023.05.004 2023

  8. [8]

    W.F.Godoy,P.Valero-Lara,K.Teranishi,P.Balaprakash,J.S.Vetter, LargeLanguageModelEvaluationforHigh-PerformanceComputing Software Development, Concurrency and Computation: Practice and Experience 36 (26) (2024) e8269.doi:10.1002/cpe.8269

    work page doi:10.1002/cpe.8269 2024

  9. [9]

    Bungartz, C

    H.-J. Bungartz, C. Riesinger, M. Schreiber, G. Snelting, A. Zwinkau, Invasive computing in HPC with X10, in: Proceedings of the third ACM SIGPLAN X10 Workshop, X10 ’13, ACM, New York, NY, USA, 2013, pp. 12–19.doi:10.1145/2481268.2481274

    work page doi:10.1145/2481268.2481274 2013

  10. [10]

    Garcia, J

    M. Garcia, J. Labarta, J. Corbalan, Hints to improve automatic load balancing with LeWI for hybrid applications, Journal of Parallel and Distributed Computing 74 (9) (2014) 2781–2794.doi:10.1016/j.jp dc.2014.05.004

    work page doi:10.1016/j.jp 2014

  11. [11]

    Lopez, J

    V. Lopez, J. Criado, R. Peñacoba, R. Ferrer, X. Teruel, M. Garcia- Gasulla, An OpenMP Free Agent Threads Implementation, in: OpenMP: Enabling Massive Node-Level Parallelism: 17th Interna- tionalWorkshoponOpenMP,IWOMP2021,Bristol,UK,September 14–16,2021,Proceedings,Springer-Verlag,Berlin,Heidelberg,2021, pp. 211–225.doi:10.1007/978-3-030-85262-7_15

    work page doi:10.1007/978-3-030-85262-7_15 2021

  12. [12]

    Huber, M

    D. Huber, M. Streubel, I. Comprés, M. Schulz, M. Schreiber, H. Pritchard, Towards dynamic resource management with mpi ses- sions and pmix, in: Proceedings of the 29th European MPI Users’ Group Meeting, EuroMPI/USA ’22, Association for Computing Ma- chinery, New York, NY, USA, 2022, p. 57–67.doi:10.1145/3555819. 3555856

    work page doi:10.1145/3555819 2022

  13. [13]

    Iserte Agut, High-throughput Computation through Efficient Re- source Management, Ph.D

    S. Iserte Agut, High-throughput Computation through Efficient Re- source Management, Ph.D. Thesis, Universitat Jaume I, accepted: 2018-12-05T10:00:25Z Publication Title: TDX (Tesis Doctorals en Xarxa) (Nov. 2018).doi:10.6035/14101.2018.176272

    work page doi:10.6035/14101.2018.176272 2018

  14. [14]

    Martín, M.-C

    G. Martín, M.-C. Marinescu, D. E. Singh, J. Carretero, FLEX-MPI: an MPI Extension for Supporting Dynamic Load Balancing on Het- erogeneousNon-dedicatedSystems,in:Euro-ParParallelProcessing, 2013, pp. 138–149

    2013

  15. [15]

    Bhattarai, H

    R. Bhattarai, H. Pritchard, S. Ghafoor, Dynamic Resource Manage- ment for Elastic Scientific Workflows using PMIx, in: 2024 IEEE S. Iserte et al.:Preprint submitted to ElsevierPage 17 of 22 MPI Malleability Validation under HPC Conditions International Parallel and Distributed Processing Symposium Work- shops (IPDPSW), 2024.doi:10.1109/IPDPSW63119.2024.00131

    work page doi:10.1109/ipdpsw63119.2024.00131 2024

  16. [16]

    S.Iserte,I.Martín-Álvarez,K.Rojek,J.I.Aliaga,M.Castillo,W.Fol- warska, A. J. Peña, Resource optimization with MPI process mal- leability for dynamic workloads in HPC clusters, Future Generation ComputerSystems(2025)107949doi:10.1016/j.future.2025.107949

    work page doi:10.1016/j.future.2025.107949 2025

  17. [17]

    ReSHAPE: A Framework for Dynamic Resizing and Scheduling of Homogeneous Applications in aParallelEnvironment,in:2007InternationalConferenceonParallel Processing (ICPP 2007), pp

    R. Sudarsan, C. J. Ribbens, ReSHAPE: A Framework for Dynamic Resizing and Scheduling of Homogeneous Applications in a Parallel Environment, in: 2007 International Conference on Parallel Process- ing (ICPP 2007), 2007, pp. 44–44.doi:10.1109/ICPP.2007.73

    work page doi:10.1109/icpp.2007.73 2007

  18. [18]

    Sarood, A

    O. Sarood, A. Langer, A. Gupta, L. Kale, Maximizing throughput of overprovisioned HPC data centers under a strict power budget, in: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’14, IEEE Press, NewOrleans,Louisana,2014,pp.807–818.doi:10.1109/SC.2014.71

    work page doi:10.1109/sc.2014.71 2014

  19. [19]

    Prabhakaran, M

    S. Prabhakaran, M. Neumann, S. Rinke, F. Wolf, A. Gupta, L. V. Kale, A Batch System with Efficient Adaptive Scheduling for Mal- leable and Evolving Applications, in: Proceedings of the 2015 IEEE InternationalParallelandDistributedProcessingSymposium,IPDPS ’15,IEEEComputerSociety,USA,2015,pp.429–438.doi:10.1109/ IPDPS.2015.34

    2015

  20. [20]

    Consciousness and Cognition20(2011) https://doi.org/10.1016/j

    S. Iserte, R. Mayo, E. S. Quintana-Ortí, V. Beltran, A. J. Peña, DMR API: Improving Cluster Productivity by Turning Applications into Malleable, Parallel Computing 78 (2018) 54–66.doi:10.1016/j. parco.2018.07.006

    work page doi:10.1016/j 2018

  21. [21]

    Sudarsan, C

    R. Sudarsan, C. J. Ribbens, D. Farkas, Dynamic Resizing of Parallel Scientific Simulations: A Case Study Using LAMMPS, in: Proceed- ings of the 9th International Conference on Computational Science: PartI,ICCS’09,Springer-Verlag,Berlin,Heidelberg,2009,pp.175– 184.doi:10.1007/978-3-642-01970-8_18

    work page doi:10.1007/978-3-642-01970-8_18 2009

  22. [22]

    A Study of the Effect of Process MalleabilityintheEnergyEfficiencyonGPU-basedClusters

    S. Iserte, K. Rojek, A Study of the Effect of Process Malleability in the Energy Efficiency on GPU-based Clusters, Journal of Supercom- puting 76 (2020) 255–274.doi:10.1007/s11227-019-03034-x

    work page doi:10.1007/s11227-019-03034-x 2020

  23. [23]

    Dynamic reconfiguration of noniterative scientific applications: A case study with HPG aligner

    S. Iserte, H. Martínez, S. Barrachina, M. Castillo, R. Mayo, A. J. Peña,Dynamicreconfigurationofnoniterativescientificapplications: A case study with HPG aligner, The International Journal of High Performance Computing Applications 33 (5) (2019) 804–816.doi: 10.1177/1094342018802347

    work page doi:10.1177/1094342018802347 2019

  24. [24]

    Zakay, D

    N. Zakay, D. G. Feitelson, Preserving User Behavior Characteristics inTrace-BasedSimulationofParallelJobScheduling,in:2014IEEE 22ndInternationalSymposiumonModelling,Analysis&Simulation ofComputerandTelecommunicationSystems,2014,pp.51–60.doi: 10.1109/MASCOTS.2014.15

    work page doi:10.1109/mascots.2014.15 2014

  25. [25]

    S. Schlagkamp, Influence of Dynamic Think Times on Parallel Job Scheduler Performances in Generative Simulations, in: 3rd confer- ence on Networked Systems Design & Implementation, 2017.doi: 10.1007/978-3-319-61756-5_7

    work page doi:10.1007/978-3-319-61756-5_7 2017

  26. [26]

    B.Schroeder,A.Wierman,M.Harchol-Balter,Openversusclosed:a cautionary tale, in: Proceedings of the 3rd conference on Networked Systems Design & Implementation - Volume 3, NSDI’06, USENIX Association, USA, 2006, p. 18

    2006

  27. [27]

    N.Zakay,D.G.Feitelson,OnIdentifyingUserSessionBoundariesin Parallel Workload Logs, in: Workshop on Job Scheduling Strategies for Parallel Processing, 2013.doi:10.1007/978-3-642-35867-8_12

    work page doi:10.1007/978-3-642-35867-8_12 2013

  28. [28]

    Madon, G

    M. Madon, G. Da Costa, J.-M. Pierson, Replay with Feedback: : How does the performance of HPC system impact user submission behavior?, Future Generation Computer Systems 155 (C) (2024) 66– 79.doi:10.1016/j.future.2024.01.024

    work page doi:10.1016/j.future.2024.01.024 2024

  29. [29]

    D. G. Feitelson, Resampling with Feedback: A New Paradigm of Using Workload Data for Performance Evaluation, in: Workshop on Job Scheduling Strategies for Parallel Processing, 2021.doi:10.100 7/978-3-030-88224-2_1

    2021

  30. [30]

    D. G. Feitelson, Packing Schemes for Gang Scheduling, in: Pro- ceedings of the Workshop on Job Scheduling Strategies for Parallel Processing, IPPS ’96, Springer-Verlag, Berlin, Heidelberg, 1996, pp. 89–110

    1996

  31. [31]

    U.Lublin,D.G.Feitelson,Theworkloadonparallelsupercomputers: modeling the characteristics of rigid jobs, Journal of Parallel and Distributed Computing 63 (11) (2003) 1105–1122.doi:10.1016/S074 3-7315(03)00108-4

    work page doi:10.1016/s074 2003

  32. [32]

    J. I. Aliaga, M. Castillo, S. Iserte, I. Martín-Álvarez, R. Mayo, A Survey on Malleability Solutions for High-Performance Distributed Computing, Applied Science 12 (2022) 1–32.doi:10.3390/app12105 231

    work page doi:10.3390/app12105 2022

  33. [33]

    A.Tarraf,M.Schreiber,A.Cascajo,J.-B.Besnard,M.-A.Vef,D.Hu- ber, S. Happ, A. Brinkmann, D. E. Singh, H.-C. Hoppe, A. Miranda, A. J. Peña, R. Machado, M. G. Gasulla, M. Schulz, P. Carpenter, S. Pickartz, T. Rotaru, S. Iserte, V. Lopez, J. Ejarque, H. Sirwani, F.Wolf,MalleabilityinModernHPCSystems:CurrentExperiences, Challenges, and Future Opportunities, IEE...

    work page doi:10.1109/tpds.2024.3406764 2024

  34. [34]

    Iserte, R

    S. Iserte, R. Mayo, E. S. Quintana-Ortí, A. J. Peña, DMRlib: Easy- coding and Efficient Resource Management for Job Malleability, IEEE Transactions on Computers 70 (2020) 1443–1457.doi:10.1 109/TC.2020.3022933

    work page arXiv 2020

  35. [35]

    D.G.Feitelson,L.Rudolph,TowardsConvergenceinJobSchedulers forParallelSupercomputers,in:ProceedingsoftheWorkshoponJob Scheduling Strategies for Parallel Processing, IPPS ’96, Springer- Verlag, Berlin, Heidelberg, 1996, pp. 1–26

    1996

  36. [36]

    S.Iserte,G.Houzeaux,P.Sandås,A.J.Peña,M.Garcia-Gasulla,Mal- leable Computational Fluid Dynamics Simulations, in: Proceedings of the 36th Parallel CFD International Conference, Merida, Yucatan, Mexico, 2025

    2025

  37. [37]

    2024).doi:10.1007/s11227-024 -06277-5

    I.Martín-Álvarez,J.I.Aliaga,M.Castillo,S.Iserte,Proteo:aframe- workforthegenerationandevaluationofmalleableMPIapplications, The Journal of Supercomputing (Jul. 2024).doi:10.1007/s11227-024 -06277-5

    work page doi:10.1007/s11227-024 2024

  38. [38]

    D.Huber,S.Iserte,M.Schreiber,A.J.Peña,M.Schulz,Bridgingthe GapBetweenGenericityandProgrammabilityofDynamicResources inHPC,in:ISCHighPerformance2025ResearchPaperProceedings (40th International Conference), 2025, pp. 1–11

    2025

  39. [39]

    Iserte, I

    S. Iserte, I. Martín-Álvarez, K. Rojek, J. I. Aliaga, M. Castillo, A. J. Peña, Towards the Democratization and Standardization of Dynamic Resources with MPI Spawning, in: R. Wyrzykowski, J. Dongarra, E. Deelman, K. Karczewski (Eds.), Parallel Processing and Applied Mathematics, Springer Nature Switzerland, Cham, 2025, pp. 287– 300.doi:10.1007/978-3-031-85697-6_19

    work page doi:10.1007/978-3-031-85697-6_19 2025

  40. [40]

    In: Proceedings of PER- MAVOST (2021)

    V. Lopez, G. Ramirez Miranda, M. Garcia-Gasulla, TALP: A Lightweight Tool to Unveil Parallel Efficiency of Large-scale Exe- cutions, in: Proceedings of the 2021 on Performance EngineeRing, Modelling,Analysis,andVisualizatiOnSTrategy,PERMAVOST’21, Association for Computing Machinery, New York, NY, USA, 2021, pp. 3–10.doi:10.1145/3452412.3462753

    work page doi:10.1145/3452412.3462753 2021

  41. [41]

    succesful execution

    K. Rojek, , R. Wyrzykowski, Parallelization of 3D MPDATA Al- gorithm Using Many Graphics Processors, in: Proceedings of the 13th International Conference on Parallel Computing Technologies - Volume 9251, Guide Proceedings, 2015, pp. 445–457.doi:10.100 7/978-3-319-21909-7_43. S. Iserte et al.:Preprint submitted to ElsevierPage 18 of 22 Appendix: Artifact D...

    2015