pith. sign in

arxiv: 2606.22175 · v1 · pith:6A2YA4TNnew · submitted 2026-06-20 · 💻 cs.DC

StickyInvoc: Rethinking Task Models for High-throughput Workflows in the LLM Era

Thanh Son Phung , Douglas Thain This is my paper

Pith reviewed 2026-06-26 11:14 UTC · model grok-4.3

classification 💻 cs.DC
keywords high-throughput computingLLM workflowstask modelssticky tasksinvocation tasksstate persistenceHPC clustersworkflow optimization
0
0 comments X

The pith

StickyInvoc decouples state creation from computation so invocation tasks inherit persistent LLM models without repeated loading.

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

The paper argues that traditional task models force every LLM inference to rebuild multi-gigabyte model states from scratch, creating prohibitive overhead at the scale of thousands of tasks on heterogeneous HPC resources. StickyInvoc introduces sticky tasks that build the state once from a template and invocation tasks that inherit it for actual work, leaving the state intact on exit. This separation amortizes the dominant setup cost across many inferences. A sympathetic reader would care because the approach turns otherwise idle GPUs into usable capacity for large scientific workflows that currently stall on repeated model transfers.

Core claim

StickyInvoc establishes a symbiotic relationship between two new task models: a sticky task creates a persistent computational state on a compute node from a user-provided template without performing goodput computation, while subsequent invocation tasks inherit that state to execute the actual computation without incurring creation or destruction overhead. The model therefore decouples the creation and destruction of computational states, allowing the state of LLM models to be created once per sticky task and amortized over many invocation tasks.

What carries the argument

The StickyInvoc paradigm of sticky tasks that create persistent state and invocation tasks that inherit it without recreation or destruction.

If this is right

  • A claim verification workflow of 150k inferences achieves a 3.6x speedup on a 20-GPU testbed.
  • The same workflow completes in 784 seconds when incrementally scaled to 186 otherwise idle GPUs.
  • Model loading costs are incurred only once per sticky task instead of once per inference task.
  • High-throughput workflows can utilize heterogeneous and preemptible resources more effectively by leaving model state resident.

Where Pith is reading between the lines

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

  • The model could apply to other state-heavy scientific codes that currently repeat large data loads, such as molecular dynamics or climate simulations.
  • Incremental scaling to idle GPUs implies the approach may improve overall cluster utilization when combined with existing schedulers.
  • Repeated state transfers avoided by this method could lower total energy use for LLM inference campaigns.
  • Explicit tests of inheritance under high preemption rates would be needed to confirm robustness beyond the reported stable testbed.

Load-bearing premise

Invocation tasks can reliably inherit and use the persistent state created by a sticky task without additional overhead or failure under the heterogeneous and preemptible conditions typical of high-throughput resources.

What would settle it

Observing that state inheritance on preemptible nodes either fails or adds measurable overhead that erases the reported 3.6x speedup on the 150k-inference workflow would disprove the central performance claim.

Figures

Figures reproduced from arXiv: 2606.22175 by Douglas Thain, Thanh Son Phung.

Figure 1
Figure 1. Figure 1: Model Loading and Inference Times across 5 Different [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Persisted States in HPC Clusters. (Left) A cluster manager first prioritizes resource allocations to static jobs on standard batch queues. (Middle) Preemptible resources allow faster access to transiently available resources, but incurs a high startup penalty from LLM model initialization. (Right) Sticky tasks create persistent LLM states on preemptible re￾sources such that invocation tasks can be readily … view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the Parsl-TaskVine Framework. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Code Example of an LLM-integrated Claim Verification [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Implementation Overview of StickyInvoc. The TaskVine scheduler analyzes F for its template upon the first invocation request, and sends it to the worker. The sticky task produces the StateManager process in the worker, which registers F’s code and creates F’s state from the template and persists it locally. This state and registered code are then used to execute the current invocation tasks, and subsequent… view at source ↗
Figure 7
Figure 7. Figure 7: Execution Time of the PromptVerify Workflow between Three Implementation Versions on Static Resources. The work￾flow is run with 3 different implementation versions: create￾destroy (no state template is encoded in the workflow, forcing an LLM state initialization per task), StickyI/O (LLM state is persisted only on local disk of remote nodes, which includes GBs of model parameters and software dependencies… view at source ↗
Figure 9
Figure 9. Figure 9: Breakdown of Inference Task Runtimes between 3 Implementation Versions (the top row shows results from tasks run [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Effect of Inference Batch Size to the Workflow’s [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Workflow Resilience Against Dynamic Availability of High-throughput Resources. [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
read the original abstract

The integration of LLMs into high-throughput workflows is creating a new class of workloads on HPC clusters that promises to accelerate advances in scientific discovery with unprecedented generative capabilities. However, the traditional task model imposes a prohibitive overhead in this new domain: each task must create its computational state from scratch and destroy it upon completion. For each LLM inference task, this "create-destroy" model forces the repeated and costly transfer of multi-gigabyte model parameters from a long-term, reliable storage to a compute node's local disk, its CPU memory, and finally its GPU memory. This overhead, compounded by the inherently high startup cost of LLM inference, the typical scale of thousands of tasks in high-throughput workflows, and the heterogeneous and preemptible nature of high-throughput resources, presents a significant performance barrier. To overcome this barrier, this paper presents StickyInvoc: a symbiotic relationship between two new task models for high-throughput workflows. Specifically, a "sticky" task creates a persistent state on a compute node from a user-provided template, but doesn't execute any goodput computation by itself. Instead, this state is then inherited by subsequent "invocation" tasks, which perform the actual computation without incurring the state creation overhead or destroying the state upon exit. StickyInvoc thus allows the decoupling of the creation and destruction of computational states, allowing the computational state of LLM models to be created once per sticky task and its cost amortized over many subsequent invocation tasks. Our evaluation shows that when rewritten in the StickyInvoc paradigm, a claim verification workflow consisting of 150k inferences achieves a 3.6x speedup on a stable testbed with 20 GPUs, and completes in just 784 seconds by incrementally scaling out to 186 otherwise idle GPUs.

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

1 major / 1 minor

Summary. The paper proposes StickyInvoc, a new task model pairing 'sticky' tasks (which create persistent computational state such as loaded LLM models from a user template without performing goodput work) with 'invocation' tasks (which inherit that state to execute inferences). It claims that rewriting a claim-verification workflow of 150k inferences under this model yields a 3.6x speedup on a stable 20-GPU testbed and finishes in 784 seconds when scaling out to 186 otherwise-idle GPUs.

Significance. If the inheritance mechanism proves reliable, the approach would amortize multi-gigabyte model-loading costs across many tasks and could materially improve throughput for LLM-driven scientific workflows on HPC clusters; the empirical scaling result on otherwise-idle GPUs is a concrete, falsifiable data point that strengthens the practical case.

major comments (1)
  1. [Abstract / Evaluation description] The abstract states that heterogeneous and preemptible resources are a core source of create-destroy overhead that StickyInvoc must overcome, yet the reported 3.6x speedup and 784-second completion are measured exclusively on a stable testbed with 20 GPUs plus otherwise-idle GPUs. No results are given for state inheritance, recovery from preemption, or cross-node heterogeneity, leaving the load-bearing assumption that invocation tasks can reliably use sticky state without added overhead or failure under the stated conditions untested.
minor comments (1)
  1. [Abstract] The abstract reports a 3.6x speedup without error bars, workload parameter details, or a description of how the baseline task model was implemented, making it difficult to judge whether post-hoc selection or unaccounted overheads affect the result.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive assessment of significance and for the detailed comment. We respond point-by-point below.

read point-by-point responses

  1. Referee: [Abstract / Evaluation description] The abstract states that heterogeneous and preemptible resources are a core source of create-destroy overhead that StickyInvoc must overcome, yet the reported 3.6x speedup and 784-second completion are measured exclusively on a stable testbed with 20 GPUs plus otherwise-idle GPUs. No results are given for state inheritance, recovery from preemption, or cross-node heterogeneity, leaving the load-bearing assumption that invocation tasks can reliably use sticky state without added overhead or failure under the stated conditions untested.

    Authors: The referee correctly notes that the reported 3.6x speedup and 784-second scaling result were obtained on a stable 20-GPU testbed plus otherwise-idle GPUs, without dedicated experiments measuring recovery from preemption or behavior under cross-node heterogeneity. State inheritance itself is directly exercised and measured by the 150k-inference claim-verification workflow that produces the 3.6x improvement. The scaling experiment further shows that invocation tasks can attach to sticky state across a larger set of idle GPUs. We agree, however, that the abstract's emphasis on heterogeneous and preemptible resources as a primary motivation is not matched by results that explicitly test those conditions. We will therefore revise the abstract and evaluation description to more precisely delineate the tested scenarios and add an explicit limitations paragraph on preemption and heterogeneity. This is a partial revision; new experiments on preemptible resources are outside the scope of the current testbed. revision: partial

Circularity Check

0 steps flagged

No circularity: speedup is empirical measurement, not derived quantity

full rationale

The paper introduces StickyInvoc as a new task model that decouples state creation (sticky tasks) from computation (invocation tasks) to amortize LLM model loading costs. Its central performance claim is an empirical observation: a claim verification workflow with 150k inferences achieves 3.6x speedup on a 20-GPU stable testbed and completes in 784s when scaling to 186 GPUs. This is presented as a direct experimental result rather than a prediction or first-principles derivation from equations or fitted parameters. No self-citations, uniqueness theorems, or ansatzes are used to support the result; the evaluation stands as an independent measurement of the implemented system. The derivation chain consists of model description followed by benchmark data, with no step reducing a claimed output to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 2 invented entities

The central claim rests on the engineering premise that state inheritance can be implemented reliably in existing HPC schedulers; no free parameters, mathematical axioms, or new physical entities are introduced in the abstract.

invented entities (2)
  • sticky task no independent evidence
    purpose: Creates and holds persistent computational state (e.g., loaded LLM) without performing user computation
    New task type defined to decouple state creation from execution
  • invocation task no independent evidence
    purpose: Performs actual LLM inference by inheriting state from a prior sticky task
    New task type defined to perform computation without state creation overhead

pith-pipeline@v0.9.1-grok · 5849 in / 1343 out tokens · 20432 ms · 2026-06-26T11:14:53.941646+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

81 extracted references · 12 canonical work pages · 3 internal anchors

  1. [1]

    GPT-4 Technical Report

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

    G. Team, P. Georgiev, V . I. Lei, R. Burnell, L. Bai, A. Gulati, G. Tanzer, D. Vincent, Z. Pan, S. Wanget al., “Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context,”arXiv preprint arXiv:2403.05530, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    The Claude 3 Model Family: Opus, Sonnet, Haiku,

    Anthropic, “The Claude 3 Model Family: Opus, Sonnet, Haiku,” 2024, available at https://www- cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model Card Claude 3.pdf

    2024

  4. [4]

    Minifold: Simple, fast, and accurate protein structure prediction,

    J. Wohlwend, M. Reveiz, M. McPartlon, A. Feldmann, W. Jin, and R. Barzilay, “Minifold: Simple, fast, and accurate protein structure prediction,”Transactions on Machine Learning Research, 2025

    2025

  5. [5]

    Energy efficient protein language models: Leveraging small language models with lora for controllable protein generation,

    A. Shah and S. Jayaratnam, “Energy efficient protein language models: Leveraging small language models with lora for controllable protein generation,”arXiv preprint arXiv:2411.05966, 2024

    work page arXiv 2024

  6. [6]

    Scaling down for effi- ciency: Medium-sized transformer models for protein sequence transfer learning,

    L. C. Vieira, M. L. Handojo, and C. O. Wilke, “Scaling down for effi- ciency: Medium-sized transformer models for protein sequence transfer learning,”bioRxiv, pp. 2024–11, 2024

    2024

  7. [7]

    Col- mena: Scalable machine-learning-based steering of ensemble simula- tions for high performance computing,

    L. Ward, G. Sivaraman, J. G. Pauloski, Y . Babuji, R. Chard, N. Dandu, P. C. Redfern, R. S. Assary, K. Chard, L. A. Curtisset al., “Col- mena: Scalable machine-learning-based steering of ensemble simula- tions for high performance computing,” in2021 IEEE/ACM Workshop on Machine Learning in High Performance Computing Environments (MLHPC). IEEE, 2021, pp. 9–20

    2021

  8. [8]

    Workflowllm: Enhancing workflow orchestration capability of large language models,

    S. Fan, X. Cong, Y . Fu, Z. Zhang, S. Zhang, Y . Liu, Y . Wu, Y . Lin, Z. Liu, and M. Sun, “Workflowllm: Enhancing workflow orchestration capability of large language models,”arXiv preprint arXiv:2411.05451, 2024

    work page arXiv 2024

  9. [9]

    A strategic coordination framework of small llms matches large llms in data synthesis,

    X. Gao, Q. Pei, Z. Tang, Y . Li, H. Lin, J. Wu, L. Wu, and C. He, “A strategic coordination framework of small llms matches large llms in data synthesis,”arXiv preprint arXiv:2504.12322, 2025

    work page arXiv 2025

  10. [10]

    Pegasus, a work- flow management system for science automation,

    E. Deelman, K. Vahi, G. Juve, M. Rynge, S. Callaghan, P. J. Maechling, R. Mayani, W. Chen, R. F. Da Silva, M. Livnyet al., “Pegasus, a work- flow management system for science automation,”Future Generation Computer Systems, vol. 46, pp. 17–35, 2015

    2015

  11. [12]

    Nextflow enables reproducible computational work- flows,

    P. Di Tommaso, M. Chatzou, E. W. Floden, P. P. Barja, E. Palumbo, and C. Notredame, “Nextflow enables reproducible computational work- flows,”Nature biotechnology, vol. 35, no. 4, pp. 316–319, 2017

    2017

  12. [13]

    Work queue+ python: A framework for scalable scientific ensemble appli- cations,

    P. Bui, D. Rajan, B. Abdul-Wahid, J. Izaguirre, and D. Thain, “Work queue+ python: A framework for scalable scientific ensemble appli- cations,” inWorkshop on python for high performance and scientific computing at sc11, 2011

    2011

  13. [14]

    Reproducible, scalable, and share- able analysis pipelines with bioinformatics workflow managers,

    L. Wratten, A. Wilm, and J. G ¨oke, “Reproducible, scalable, and share- able analysis pipelines with bioinformatics workflow managers,”Nature methods, vol. 18, no. 10, pp. 1161–1168, 2021

    2021

  14. [15]

    The nf-core framework for community-curated bioinformatics pipelines,

    P. A. Ewels, A. Peltzer, S. Fillinger, H. Patel, J. Alneberg, A. Wilm, M. U. Garcia, P. Di Tommaso, and S. Nahnsen, “The nf-core framework for community-curated bioinformatics pipelines,”Nature biotechnology, vol. 38, no. 3, pp. 276–278, 2020

    2020

  15. [16]

    A scalable scenic workflow for single-cell gene regulatory network analysis,

    B. Van de Sande, C. Flerin, K. Davie, M. De Waegeneer, G. Hulselmans, S. Aibar, R. Seurinck, W. Saelens, R. Cannoodt, Q. Rouchonet al., “A scalable scenic workflow for single-cell gene regulatory network analysis,”Nature protocols, vol. 15, no. 7, pp. 2247–2276, 2020

    2020

  16. [17]

    Reshaping high en- ergy physics applications for near-interactive execution using taskvine,

    B. Sly-Delgado, B. Tovar, J. Zhou, and D. Thain, “Reshaping high en- ergy physics applications for near-interactive execution using taskvine,” inSC24: International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE, 2024, pp. 1–13

    2024

  17. [18]

    Dynamic task shaping for high throughput data analysis applications in high energy physics,

    B. Tovar, B. Lyons, K. Mohrman, B. Sly-Delgado, K. Lannon, and D. Thain, “Dynamic task shaping for high throughput data analysis applications in high energy physics,” in2022 IEEE International Parallel and Distributed Processing Symposium (IPDPS). IEEE, 2022, pp. 346– 356

    2022

  18. [19]

    Pegasus workflow management system: helping applications from earth and space,

    G. Mehta, E. Deelman, K. Vahi, and F. Silva, “Pegasus workflow management system: helping applications from earth and space,” inAGU Fall Meeting Abstracts, vol. 2010, 2010, pp. IN41B–1362

    2010

  19. [20]

    Gemma 2b,

    Google, “Gemma 2b,” Hugging Face, 2025, accessed: 2025-10-01. [Online]. Available: https://huggingface.co/google/gemma-2-2b

    2025

  20. [21]

    Smollm2-1.7b-instruct,

    H. Face, “Smollm2-1.7b-instruct,” Hugging Face, 2025, accessed: 2025-10-01. [Online]. Available: https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct

    2025

  21. [22]

    Stablelm-2-1.6b,

    S. AI, “Stablelm-2-1.6b,” Hugging Face, 2025, accessed: 2025-10-01. [Online]. Available: https://huggingface.co/stabilityai/stablelm-2-1 6b

    2025

  22. [23]

    Qwen2.5-1.5b,

    U. AI, “Qwen2.5-1.5b,” Hugging Face, 2025, accessed: 2025-10-01. [Online]. Available: https://huggingface.co/unsloth/Qwen2.5-1.5B

    2025

  23. [24]

    Deepseek-r1-distill-qwen-1.5b,

    DeepSeek-AI, “Deepseek-r1-distill-qwen-1.5b,” Hugging Face, 2025, accessed: 2025-10-01. [Online]. Available: https://huggingface.co/deepseek- ai/DeepSeek-R1-Distill-Qwen-1.5B

    2025

  24. [25]

    Ohio Supercomputer Center

    (2025) Monitoring and managing your job. Ohio Supercomputer Center. Accessed: 2025-10-01. [Online]. Available: https://www.osc.edu/ supercomputing/batch-processing-at-osc/monitoring-and-managing- your-job

    2025

  25. [26]

    Princeton Research Computing

    (2025) Job priority. Princeton Research Computing. Accessed: 2025- 10-01. [Online]. Available: https://researchcomputing.princeton.edu/ support/knowledge-base/job-priority

    2025

  26. [27]

    Argonne Leadership Computing Facility (ALCF)

    (2025) Queue scheduling. Argonne Leadership Computing Facility (ALCF). Accessed: 2025-10-01. [Online]. Available: https://docs.alcf.anl.gov/policies/queue-scheduling/

    2025

  27. [28]

    2025 global semiconductor industry outlook,

    Deloitte Insights, “2025 global semiconductor industry outlook,” February 2025. [Online]. Available: https://www.deloitte.com/us/en/insights/industry/technology/technology- media-telecom-outlooks/semiconductor-industry-outlook.html

    2025

  28. [29]

    Trends in ai supercomputers,

    K. F. Pilz, J. Sanders, R. Rahman, and L. Heim, “Trends in ai supercomputers,”arXiv preprint arXiv:2504.16026, 2025

    work page arXiv 2025

  29. [30]

    A survey on hardware accelerators for large language models,

    C. Kachris, “A survey on hardware accelerators for large language models,”Applied Sciences, vol. 15, no. 2, p. 586, 2025

    2025

  30. [31]

    Scheduling deep learning jobs in multi-tenant gpu clusters via wise resource sharing,

    Y . Luo, Q. Wang, S. Shi, J. Lai, S. Qi, J. Zhang, and X. Wang, “Scheduling deep learning jobs in multi-tenant gpu clusters via wise resource sharing,” in2024 IEEE/ACM 32nd International Symposium on Quality of Service (IWQoS). IEEE, 2024, pp. 1–10

    2024

  31. [32]

    Resource allocation and workload scheduling for large-scale distributed deep learning: A survey,

    F. Liang, Z. Zhang, H. Lu, C. Li, V . Leung, Y . Guo, and X. Hu, “Resource allocation and workload scheduling for large-scale distributed deep learning: A survey,”arXiv preprint arXiv:2406.08115, 2024

    work page arXiv 2024

  32. [33]

    Mirage: Towards low-interruption services on batch gpu clusters with reinforce- ment learning,

    Q. Ding, P. Zheng, S. Kudari, S. Venkataraman, and Z. Zhang, “Mirage: Towards low-interruption services on batch gpu clusters with reinforce- ment learning,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 2023, pp. 1–13

    2023

  33. [34]

    Multi-tenant gpu clusters for deep learning workloads: Anal- ysis and implications,

    M. Jeon, S. Venkataraman, J. Qian, A. Phanishayee, W. Xiao, and F. Yang, “Multi-tenant gpu clusters for deep learning workloads: Anal- ysis and implications,”Technical report, Microsoft Research, 2018

    2018

  34. [35]

    {AntMan}: Dynamic scaling on{GPU}clusters for deep learning,

    W. Xiao, S. Ren, Y . Li, Y . Zhang, P. Hou, Z. Li, Y . Feng, W. Lin, and Y . Jia, “{AntMan}: Dynamic scaling on{GPU}clusters for deep learning,” in14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), 2020, pp. 533–548

    2020

  35. [36]

    Analysis of{Large-Scale}{Multi-Tenant}{GPU}clusters for{DNN}training workloads,

    M. Jeon, S. Venkataraman, A. Phanishayee, J. Qian, W. Xiao, and F. Yang, “Analysis of{Large-Scale}{Multi-Tenant}{GPU}clusters for{DNN}training workloads,” in2019 USENIX Annual Technical Conference (USENIX ATC 19), 2019, pp. 947–960

    2019

  36. [37]

    (2025) Queues and charges. NERSC. Accessed: 2025-10-02. [Online]. Available: https://docs.nersc.gov/jobs/policy/#qos-cost-factor- charge-multipliers-and-discounts

    2025

  37. [38]

    NCAR HPC

    (2025) Job premption with pbs. NCAR HPC. Ac- cessed: 2025-10-02. [Online]. Available: https://ncar-hpc- docs.readthedocs.io/en/latest/pbs/preemption/#charging-and-allocations

    2025

  38. [39]

    University of Maryland High- Performance Computing Center

    (2025) Available hpc partitions. University of Maryland High- Performance Computing Center. Accessed: 2025-10-02. [Online]. Available: https://hpcc.umd.edu/kb/queues/#scavenger-partition

    2025

  39. [40]

    Center for High Performance Computing at the University of Utah

    (2025) Atomatic restarting of preemptable jobs. Center for High Performance Computing at the University of Utah. Accessed: 2025-10-02. [Online]. Available: https://www.chpc.utah.edu/documentation/software/slurm-job- preemption.php#Automatic%20Restarting%20of%20Preemptable%20Jobs

    2025

  40. [41]

    Fermilab

    (2025) Slurm job scheduler. Fermilab. Accessed: 2025-10-02. [Online]. Available: https://computing.fnal.gov/wilsoncluster/slurm-job-scheduler/

    2025

  41. [42]

    Center for Computational Research at the University at Buffalo

    (2025) Slurm directives, partitions & qos. Center for Computational Research at the University at Buffalo. Accessed: 2025-10-02. [On- line]. Available: https://docs.ccr.buffalo.edu/en/latest/hpc/jobs/#slurm- directives-partitions-qos

    2025

  42. [43]

    {AW ARE}: Automate workload autoscaling with reinforcement learning in production cloud systems,

    H. Qiu, W. Mao, C. Wang, H. Franke, A. Youssef, Z. T. Kalbarczyk, T. Bas ¸ar, and R. K. Iyer, “{AW ARE}: Automate workload autoscaling with reinforcement learning in production cloud systems,” in2023 USENIX Annual Technical Conference (USENIX ATC 23), 2023, pp. 387–402

    2023

  43. [44]

    Optscaler: A collaborative framework for robust autoscaling in the cloud,

    D. Zou, W. Lu, Z. Zhu, X. Lu, J. Zhou, X. Wang, K. Liu, K. Wang, R. Sun, and H. Wang, “Optscaler: A collaborative framework for robust autoscaling in the cloud,”Proceedings of the VLDB Endowment, vol. 17, no. 12, pp. 4090–4103, 2024

    2024

  44. [45]

    Tuning a kubernetes horizontal pod autoscaler for meeting performance and load demands in cloud deployments,

    D. R. Augustyn, Ł. Wyci ´slik, and M. Sojka, “Tuning a kubernetes horizontal pod autoscaler for meeting performance and load demands in cloud deployments,”Applied Sciences, vol. 14, no. 2, p. 646, 2024

    2024

  45. [46]

    A survey on auto-scaling: how to exploit cloud elasticity,

    M. Catillo, U. Villano, and M. Rak, “A survey on auto-scaling: how to exploit cloud elasticity,”International Journal of Grid and Utility Computing, vol. 14, no. 1, pp. 37–50, 2023

    2023

  46. [47]

    Checkpointing techniques in distributed systems: A synopsis of diverse strategies over the last decades,

    H. Goulart, A. Franco, and O. Mendizabal, “Checkpointing techniques in distributed systems: A synopsis of diverse strategies over the last decades,” inWorkshop de Testes e Toler ˆancia a Falhas (WTF). SBC, 2023, pp. 15–28

    2023

  47. [48]

    Mcrengine: A scalable checkpointing system using data-aware aggregation and compression,

    T. Z. Islam, K. Mohror, S. Bagchi, A. Moody, B. R. De Supinski, and R. Eigenmann, “Mcrengine: A scalable checkpointing system using data-aware aggregation and compression,” inSC’12: Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis. IEEE, 2012, pp. 1–11

    2012

  48. [49]

    Checkmate: Evaluating checkpointing protocols for streaming dataflows,

    G. Siachamis, K. Psarakis, M. Fragkoulis, A. Van Deursen, P. Carbone, and A. Katsifodimos, “Checkmate: Evaluating checkpointing protocols for streaming dataflows,” in2024 IEEE 40th international conference on data engineering (ICDE). IEEE, 2024, pp. 4030–4043

    2024

  49. [50]

    Parsl: Per- vasive parallel programming in python,

    Y . Babuji, A. Woodard, Z. Li, D. S. Katz, B. Clifford, R. Kumar, L. Lacinski, R. Chard, J. M. Wozniak, I. Fosteret al., “Parsl: Per- vasive parallel programming in python,” inProceedings of the 28th International Symposium on High-Performance Parallel and Distributed Computing, 2019, pp. 25–36

    2019

  50. [51]

    Taskvine: Managing in-cluster storage for high- throughput data intensive workflows,

    B. Sly-Delgado, T. S. Phung, C. Thomas, D. Simonetti, A. Hennessee, B. Tovar, and D. Thain, “Taskvine: Managing in-cluster storage for high- throughput data intensive workflows,” inProceedings of the SC’23 Work- shops of the International Conference on High Performance Computing, Network, Storage, and Analysis, 2023, pp. 1978–1988

    2023

  51. [52]

    Maximizing data utility for hpc python workflow execution,

    T. S. Phung, B. Clifford, K. Chard, and D. Thain, “Maximizing data utility for hpc python workflow execution,” inProceedings of the SC’23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis, 2023, pp. 637–640

    2023

  52. [53]

    Accelerating function-centric applications by discovering, distributing, and retaining reusable context in workflow systems,

    T. S. Phung, C. Thomas, L. Ward, K. Chard, and D. Thain, “Accelerating function-centric applications by discovering, distributing, and retaining reusable context in workflow systems,” inProceedings of the 33rd International Symposium on High-Performance Parallel and Distributed Computing, 2024, pp. 122–134

    2024

  53. [54]

    Adaptive task-oriented resource allocation for large dynamic workflows on opportunistic resources,

    T. S. Phung and D. Thain, “Adaptive task-oriented resource allocation for large dynamic workflows on opportunistic resources,” in2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS). IEEE, 2024, pp. 300–311

    2024

  54. [55]

    Towards llm-based fact verification on news claims with a hierarchical step-by-step prompting method,

    X. Zhang and W. Gao, “Towards llm-based fact verification on news claims with a hierarchical step-by-step prompting method,”arXiv preprint arXiv:2310.00305, 2023

    work page arXiv 2023

  55. [56]

    Molecular facts: Desiderata for decontex- tualization in llm fact verification,

    A. Gunjal and G. Durrett, “Molecular facts: Desiderata for decontex- tualization in llm fact verification,”arXiv preprint arXiv:2406.20079, 2024

    work page arXiv 2024

  56. [57]

    FEVER: a large-scale dataset for fact extraction and VERification,

    J. Thorne, A. Vlachos, C. Christodoulopoulos, and A. Mittal, “FEVER: a large-scale dataset for fact extraction and VERification,” inNAACL- HLT, 2018

    2018

  57. [58]

    SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

    L. B. Allal, A. Lozhkov, E. Bakouch, G. M. Bl ´azquez, G. Penedo, L. Tunstall, A. Marafioti, H. Kydl ´ıˇcek, A. P. Lajar ´ın, V . Srivastav et al., “Smollm2: When smol goes big–data-centric training of a small language model,”arXiv preprint arXiv:2502.02737, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  58. [59]

    Distributed computing in practice: the condor experience,

    D. Thain, T. Tannenbaum, and M. Livny, “Distributed computing in practice: the condor experience,”Concurrency and computation: prac- tice and experience, vol. 17, no. 2-4, pp. 323–356, 2005

    2005

  59. [60]

    Taming metadata storms in parallel filesystems with metafs,

    T. Shaffer and D. Thain, “Taming metadata storms in parallel filesystems with metafs,” inProceedings of the 2nd Joint International Workshop on Parallel Data Storage & Data Intensive Scalable Computing Systems, 2017, pp. 25–30

    2017

  60. [61]

    Scalable performance of the panasas parallel file system

    B. Welch, M. Unangst, Z. Abbasi, G. A. Gibson, B. Mueller, J. Small, J. Zelenka, and B. Zhou, “Scalable performance of the panasas parallel file system.” inFAST, vol. 8, 2008, pp. 1–17

    2008

  61. [62]

    Anaconda software distribution,

    A. Inc., “Anaconda software distribution,” https://docs.anaconda.com/, 2020

    2020

  62. [63]

    Google Cloud

    (2025) Spot vms. Google Cloud. [Online]. Available: https://cloud.google.com/solutions/spot-vms

    2025

  63. [64]

    Amazon Web Services (AWS)

    (2025) Amazon ec2 spot instances. Amazon Web Services (AWS). [Online]. Available: https://aws.amazon.com/ec2/spot/

    2025

  64. [65]

    Microsoft Azure

    (2025) Spot virtual machines. Microsoft Azure. [Online]. Available: https://azure.microsoft.com/en-us/products/virtual-machines/spot

    2025

  65. [66]

    Skyserve: Serving ai models across regions and clouds with spot instances,

    Z. Mao, T. Xia, Z. Wu, W.-L. Chiang, T. Griggs, R. Bhardwaj, Z. Yang, S. Shenker, and I. Stoica, “Skyserve: Serving ai models across regions and clouds with spot instances,” inProceedings of the Twentieth European Conference on Computer Systems, 2025, pp. 159–175

    2025

  66. [67]

    Spotserve: Serving generative large language models on preemptible instances,

    X. Miao, C. Shi, J. Duan, X. Xi, D. Lin, B. Cui, and Z. Jia, “Spotserve: Serving generative large language models on preemptible instances,” in Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, 2024, pp. 1112–1127

    2024

  67. [68]

    Ama- zon Web Services (AWS)

    (2025) Spot instance interruption notices. Ama- zon Web Services (AWS). [Online]. Avail- able: https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/spot- instance-termination-notices.html

    2025

  68. [69]

    Google Cloud

    (2025) Spot vms. Google Cloud. [Online]. Available: https://cloud.google.com/compute/docs/instances/spot

    2025

  69. [70]

    Microsoft Azure

    (2025) Spot virtual machines. Microsoft Azure. [Online]. Available: https://learn.microsoft.com/en-us/azure/virtual-machines/spot-vms

    2025

  70. [71]

    Fast inference from transform- ers via speculative decoding,

    Y . Leviathan, M. Kalman, and Y . Matias, “Fast inference from transform- ers via speculative decoding,” inInternational Conference on Machine Learning. PMLR, 2023, pp. 19 274–19 286

    2023

  71. [72]

    Accelerating llm inference with staged specula- tive decoding,

    B. Spector and C. Re, “Accelerating llm inference with staged specula- tive decoding,”arXiv preprint arXiv:2308.04623, 2023

    work page arXiv 2023

  72. [73]

    Cascade speculative drafting for even faster llm inference,

    Z. Chen, X. Yang, J. Lin, C. Sun, K. Chang, and J. Huang, “Cascade speculative drafting for even faster llm inference,”Advances in Neural Information Processing Systems, vol. 37, pp. 86 226–86 242, 2024

    2024

  73. [74]

    Specexec: Massively parallel speculative decoding for interactive llm inference on consumer devices,

    R. Svirschevski, A. May, Z. Chen, B. Chen, Z. Jia, and M. Ryabinin, “Specexec: Massively parallel speculative decoding for interactive llm inference on consumer devices,”Advances in Neural Information Pro- cessing Systems, vol. 37, pp. 16 342–16 368, 2024

    2024

  74. [75]

    Efficient memory management for large language model serving with pagedattention,

    W. Kwon, Z. Li, S. Zhuang, Y . Sheng, L. Zheng, C. H. Yu, J. Gonzalez, H. Zhang, and I. Stoica, “Efficient memory management for large language model serving with pagedattention,” inProceedings of the 29th Symposium on Operating Systems Principles, 2023, pp. 611–626

    2023

  75. [76]

    Infinite-llm: Efficient llm service for long context with distattention and distributed kvcache.arXiv preprint arXiv:2401.02669, 2024

    B. Lin, C. Zhang, T. Peng, H. Zhao, W. Xiao, M. Sun, A. Liu, Z. Zhang, L. Li, X. Qiuet al., “Infinite-llm: Efficient llm service for long context with distattention and distributed kvcache,”arXiv preprint arXiv:2401.02669, 2024

    work page arXiv 2024

  76. [77]

    {ServerlessLLM}:{Low-Latency}serverless inference for large language models,

    Y . Fu, L. Xue, Y . Huang, A.-O. Brabete, D. Ustiugov, Y . Patel, and L. Mai, “{ServerlessLLM}:{Low-Latency}serverless inference for large language models,” in18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24), 2024, pp. 135–153

    2024

  77. [78]

    Middleware building blocks for workflow systems,

    M. Turilli, V . Balasubramanian, A. Merzky, I. Paraskevakos, and S. Jha, “Middleware building blocks for workflow systems,”Computing in Science & Engineering, vol. 21, no. 4, pp. 62–75, 2019

    2019

  78. [79]

    Deploying high throughput scientific workflows on container schedulers with makeflow and mesos,

    C. Zheng, B. Tovar, and D. Thain, “Deploying high throughput scientific workflows on container schedulers with makeflow and mesos,” in2017 17th IEEE/ACM International Symposium on Cluster, Cloud and Grid Computing (CCGRID). IEEE, 2017, pp. 130–139

    2017

  79. [80]

    Not all tasks are created equal: Adaptive resource allocation for heterogeneous tasks in dynamic workflows,

    T. S. Phung, L. Ward, K. Chard, and D. Thain, “Not all tasks are created equal: Adaptive resource allocation for heterogeneous tasks in dynamic workflows,” in2021 IEEE Workshop on Workflows in Support of Large- Scale Science (WORKS). IEEE, 2021, pp. 17–24

    2021

  80. [81]

    Dask: Parallel computation with blocked algorithms and task scheduling

    M. Rocklinet al., “Dask: Parallel computation with blocked algorithms and task scheduling.” inSciPy, 2015, pp. 126–132

    2015

Showing first 80 references.