Pith. sign in

REVIEW 4 major objections 6 minor 25 references

Green Distributed AI Training: Orchestrating Compute Across Renewable-Powered Micro Datacenters

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Migrating single-GPU AI training jobs between renewable-powered micro-datacenters is almost always energetically worthwhile; the binding constraint is whether the checkpoint can be transferred inside the renewable-energy window.

desk verdict A sensible feasibility formula for migrating single-GPU training jobs across renewable windows, but the quantitative claims rest on internally inconsistent numbers and an unvalidated forecast/bandwidth assumption. read the letter →

arxiv 2511.16182 v2 pith:KK4HL4EJ submitted 2025-11-20 cs.NI

Giuseppe Tomei , Andrea Mayer , Giuseppe Alcini , Stefano Salsano This is my paper
classification cs.NI
keywords greenAIrenewableenergyworkloadmigrationmicro-datacentersfeasibilityanalysischeckpointingenergy-awareschedulingjobcompletiontime
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish when it is actually practical to move a running AI training job from one small datacenter to another to chase surplus renewable power. It argues that the energy cost of migration is negligible—breakeven occurs within minutes—so the real question is time: can the checkpoint transfer, load, and downtime finish before the renewable window ends. The paper distills this into a formal feasibility condition and a three-class workload taxonomy, then shows that an orchestrator enforcing those boundaries can cut non-renewable energy use by 52% and improve average job completion time by 18% compared with static placement. A sympathetic reader would care because renewable curtailment is growing, and this offers a concrete, quantified path to turning otherwise wasted energy into useful computation.

What carries the argument

The central object is the feasibility-domain inequality T_transfer + T_load + T_downtime < α·T_energy with α = 0.1, instantiated by T_transfer = checkpoint_size / bandwidth. The paper uses this inequality to draw a phase diagram mapping checkpoint size against WAN capacity, then derives the A/B/C workload classes that act as a hard filter inside the orchestrator. The filter rejects Class C workloads and any Class B migration whose total time would exceed 10% of the remaining renewable window; optimization over renewable availability and queueing delay happens only after that filter passes.

What would settle it

A trace-driven experiment in which effective WAN bandwidth is randomly degraded by 20–50% during some migrations, or in which renewable-window forecasts carry errors matching real curtailment signals, would settle the claim: if the Feasibility-aware policy's non-renewable energy reduction drops well below 52% and its JCT improvement flips positive, then the filter's value depends on perfect inputs rather than on the feasibility condition itself.

Watch

Extended reading notes

Core claim

The paper's central claim is that for single-GPU training workloads, migration between renewable-powered micro-datacenters is energy-feasible within minutes, while time-feasibility is governed by the condition T_transfer + T_load + T_downtime < α·T_energy, with α = 0.1. Here T_energy is the remaining duration of the renewable-surplus window, and T_transfer is checkpoint size divided by wide-area bandwidth. Because a 40 GB checkpoint over 10 Gbps costs only about 0.016 kWh and reaches energy breakeven in roughly 1.3 minutes, the energy constraint virtually never binds. Instead, checkpoint size and WAN bandwidth separate workloads into Class A (transfer under 60 seconds), Class B (60–300 secon

Load-bearing premise

The whole benefit hinges on the assumption that the orchestrator's forecast of the remaining renewable window and its estimate of effective WAN bandwidth stay accurate for the full duration of each migration; the paper itself states that it assumes stable effective bandwidth and leaves forecast uncertainty unimplemented in the evaluation, so if real bandwidth or window length is worse than predicted, migrations the filter lets through will stall and the claimed 52% / 18% gain

Editorial extensions

If this is right

  • Renewable-window duration, not migration energy, sets the practical boundary: migrations are energetically justified for essentially all realistic checkpoint sizes.
  • Only checkpoints below roughly 20 GB migrate efficiently on 1–10 Gbps links; larger training states need 100 Gbps-class connectivity or techniques that shrink transfer time.
  • Purely energy-driven migration without feasibility checks increases average job completion time by about 35%, while feasibility-aware migration improves it by 18%.
  • Enforcing the feasibility filter keeps total migration overhead below 2%, since infeasible migrations are never attempted.
  • Expanding the feasible envelope—via compression, pre-staging, incremental checkpoints, or faster links—matters more than refining scheduling heuristics alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If real WAN bandwidth fluctuates during migrations, the 52% energy / 18% JCT result depends on how conservatively the orchestrator estimates bandwidth; a sensitivity test with bandwidth degradation would show how quickly the benefit erodes.
  • Operators could adopt the simple rule 'migrate only if checkpoint_size / effective_bandwidth < 0.1 × forecast renewable window, minus load and downtime' without building the full orchestration stack.
  • The paper formalizes forecast uncertainty with a risk parameter ε but never evaluates it; the next testable step is to inject realistic forecast errors into the trace and measure at what error level Feasibility-aware degrades toward Energy-only behavior.
  • Large distributed training jobs are classified as infeasible only because they must move a full checkpoint; migrating optimizer shards or compressed deltas, as the paper outlines, could make parts of those jobs feasible and turn Class C into conditional Class B.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a feasibility-domain model for migrating single-GPU AI training workloads across renewable-powered micro-datacenters. It derives a time condition (Eq. 1) and an energy-breakeven condition (Eq. 2), introduces A/B/C workload classes, and builds an orchestrator that uses the feasibility condition as a hard filter before optimizing migration targets. A 5-node trace-driven simulation is used to claim that a feasibility-aware policy reduces non-renewable energy by 52% and improves average JCT by 18% relative to static placement, while an energy-only policy degrades JCT by 35%. The paper also discusses hardware efficiency, system-level implications, and longer-term architectural and economic considerations.

Significance. The core energy-breakeven insight—that migration energy cost is negligible compared with multi-hour renewable windows, so time constraints dominate—is simple, plausible, and potentially useful for the renewable-aware scheduling community. The design of an explicit feasibility filter that prevents migrations likely to overrun their renewable window is a sensible and practically motivated idea. The paper clearly states assumptions and limitations, and the model is straightforward to reproduce from the given equations. However, the quantitative results are undermined by internal numerical contradictions (Fig. 1 vs. §IV.D, Table VII vs. §IV.D) and by a validation procedure that is largely circular. If these issues are fixed, the paper would make a modest but useful contribution; in its current form, the central quantitative claims are not established.

major comments (4)
  1. [§IV.C/E/F, Eq. (1), Fig. 2, Table IV] The time-feasibility condition Eq. (1) with α=0.1 and the stated typical window T_energy=2.5 h gives a budget T_transfer+T_load+T_downtime < 900 s. Using the paper's own Table III, a 100 GB checkpoint at 1 Gbps transfers in 800 s and a 280 GB checkpoint at 10 Gbps in 224 s, so both are time-feasible under Eq. (1). Yet Table IV and Fig. 2 label >100 GB as Class C 'Exceeds wind.,' and Table VII marks LLaMA-70B (280 GB) INFEASIBLE at 10 Gbps. The 60 s/300 s class boundaries are absolute transfer times with no dependence on T_d, although Eq. (1) makes feasibility depend jointly on T_d and bandwidth. Since the orchestrator's pruning and the Table VI gains rest on these classes, this inconsistency is load-bearing and must be resolved.
  2. [§IV.D vs. Fig. 1 and Table VII] §IV.D computes for a 40 GB checkpoint over 10 Gbps: T_transfer = 0.0089 h (32 s) and T_breakeven ≈ 1.3 min. Figure 1 lists the same 40 GB case with T_transfer = 1.28 min and T_breakeven = 3.07 min, contradicting both §IV.D and Table III (34 s). Table VII further classifies GPT-2 Medium (40 GB) as Class B and 'INFEASIBLE (Energy),' although §IV.D proves the energy breakeven is ~1.3 min, far below a 2.5 h window. Either the transfer-time model or Fig. 1/Table VII is wrong; the contradiction affects the derivation of feasibility boundaries and the classification used in the evaluation.
  3. [§VII.A and §VII.C] The validation experiments are circular. Jobs are assigned to classes using the same T_transfer = S/B model and the same 60 s/300 s cutoffs that define Class A/B/C, and the simulator's feasibility filter enforces Eq. (1). The finding that predicted classes match measured overhead is therefore built into the setup; it does not validate the model as 'predictive, and not only descriptive' (§VII.A). An independent test would run the simulator without the feasibility filter, vary B_s,d and T_d, and compare actual migration completions/failures against Eq. (1). As written, the model's predictive success is assumed rather than demonstrated.
  4. [§VI.H, §VIII.F, Table VI] The headline 52% non-renewable reduction and −18% JCT depend on the filter correctly pruning migrations that cannot finish within the renewable window. That correctness requires accurate forecasts of T_d and effective bandwidth B_s,d. §VIII.F concedes that bandwidth is assumed stable and that background traffic can push Class A jobs toward Class B; §VI.H introduces a stochastic window model with risk budget ε but never instantiates it in the evaluation. No sensitivity analysis over forecast error or bandwidth variance is reported, so the advantage over Energy-only is conditional on inputs the paper itself describes as uncertain. At minimum, the authors should report results as a function of forecast error/ε and of bandwidth degradation.
minor comments (6)
  1. [References] Reference [4] is a placeholder ('Edge AI System Vendor' with example-vendor.com) and should be replaced with a real source or removed. Table I's hardware estimates would benefit from explicit citations for each power/perf figure.
  2. [§V] Typo: 'see the the figure provided' should be 'see the figure provided'.
  3. [§VI.C] The statement 'We fix α=0.1 following the empirical thresholds used in our simulations' is circular and does not justify α. An independent rationale or a sensitivity analysis over α is needed.
  4. [Algorithm 1] Line 11 defines T_breakeven_energy = E_cost_energy / 0.75 kW; the units of E_cost_energy and energy_forecast.duration should be made explicit to avoid hour/minute confusion.
  5. [References] Formatting errors: 'Maoet al.' in [13] and 'Rajbhandari at al.' in [20] should be 'Mao et al.' and 'Rajbhandari et al.'.
  6. [§II.C] The ViT-B/32 energy-per-sample figures (2.7 mJ/sample vs. 6–7 mJ/sample) are stated without a measurement methodology or citation; please indicate whether these are new measurements, vendor estimates, or from prior work.

Circularity Check

3 steps flagged · score 6.0 of 10

Feasibility validation is closed by construction: α and the 60s/300s class cutoffs are set from the same simulation that is then used to 'confirm' the model; the 52%/−18% policy result is conditional on this hand-set filter.

  1. fitted input called prediction [§VI.C (Eq. 1), §V Algorithm 1, Table V, §VII.A]
    "We fix α=0.1 following the empirical thresholds used in our simulations."

    α is the only free parameter of the central feasibility inequality, and it is set 'following the empirical thresholds used in our simulations' (cf. Table V: 'Acceptable Overhead 10% Target'). The same α is then used both in Algorithm 1's hard filter (T_cost_time > 0.1×energy_forecast.duration) and in §VII.A, where Class A workloads are said to exhibit '<10% JCT overhead' while Class B lie 'at or above this budget'. The experiment therefore confirms the model against a budget that was used to define the model; the 'prediction' is fitted to the simulation and cannot independently validate the feasibility condition.

  2. self definitional [§VI.D (Feasibility Classes) and §VII.A (Feasibility Validation)]
    "Class A if T_transfer <60s, Class B if 60s≤T_transfer <300s, and Class C otherwise. ... assign jobs to Class A/B/C using the 60s and 300s thresholds that define the feasibility contours."

    The A/B/C classes are defined by the same T_transfer cutoffs that the validation uses to assign jobs: §VI.D sets A: T_mig<60s, B: 60≤T_mig<300s, C: T_mig≥300s, and §VII.A 'assign[s] jobs to Class A/B/C using the 60s and 300s thresholds that define the feasibility contours.' Since the simulator computes transfer time with the identical formula (T_transfer = 8S_j/B_s,d in Algorithm 1), the labels and the 'predicted bands' are the same function of the same inputs; observing that measured overhead falls into those bands is a consistency check, not an empirical confirmation.

1 more flagged steps
  1. self definitional [§VII.A (Feasibility Validation) / Table VII]
    "This one-to-one alignment between predicted T_transfer bands and measured JCT overhead confirms that the feasibility-domain model is predictive at 10 Gbps, and not only descriptive."

    The sentence claims that the model 'is predictive at 10 Gbps, and not only descriptive,' but the experiment uses the model's own thresholds to create the classes and the model's own transfer-time formula to drive the simulator, so the one-to-one alignment is guaranteed by construction rather than tested. The validation also mislabels a 40 GB checkpoint at 10 Gbps as Class B even though T_transfer ≈ 34 s < 60 s, so the asserted alignment with T_transfer bands is not even arithmetically consistent with the paper's own Eq. (1).

full rationale

The energy-breakeven analysis (Eq. 2, Fig. 1) is self-contained arithmetic from stated power and checkpoint/bandwidth values; the conclusion that energy is not the dominant constraint is an ordinary consequence of those inputs, not circular. No load-bearing self-citations or imported uniqueness theorems appear; the external references (e.g., PhoenixOS, ServerlessLLM, Liu et al.) provide independent support for the parameter values. The circularity is concentrated in the feasibility validation loop: α is fixed from the simulations ("following the empirical thresholds used in our simulations"), the 60s/300s class boundaries are the paper's own definitions, and §VII.A then assigns jobs with those same thresholds and the same transfer-time formula before declaring the model predictive. That validation is closed by construction. The Table VI policy comparison (52% non-renewable reduction, −18% JCT) is a real design evaluation, but it is conditional on the hand-set α and thresholds; it does not independently validate the feasibility-domain model. Overall, the central numerical claims do not reduce entirely to their inputs, but the paper's stated validation of the model is circular, so the score is 6 rather than higher.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The quantitative conclusions rest on a small set of hand-set parameters (α=0.1, 60/300 s cutoffs, P_sys=1.8 kW, P_node=0.75 kW, 2.5 h window) plus explicit scope assumptions (single-GPU, stable bandwidth, deterministic windows). The cited systems justify the transfer-dominance formula and its constants; they do not justify the thresholds, whose only stated basis is the simulations that then 'validate' the model. No new entities are invented.

free parameters (6)
  • α (acceptable overhead fraction) = 0.1
    Defines the entire feasibility boundary via T_transfer + T_load + T_downtime < α·T_energy. Justified in §VI.C only as 'following the empirical thresholds used in our simulations' — i.e., chosen to match the simulation that is then used to validate the model.
  • Class thresholds (A/B/C cutoffs) = 60 s / 300 s
    Hand-set boundaries in §V and §VI.D that drive workload classification, the phase diagram, and the claimed predictive accuracy. The '≤20 GB' feasibility claim is a restatement of these cutoffs, and the cutoffs conflict with Table IV's size-based classes.
  • P_sys (transfer power) = 1.8 kW
    Combined system power during transfer (§IV.D, Table V). The 'energy is negligible' conclusion is a direct consequence of this value versus P_node; stated as 'consistent with the measurements used' without a citable measurement.
  • P_node (compute power at destination) = 0.75 kW
    Sets E_benefit per hour and hence T_breakeven (§IV.D). Taken from the hardware section's unverified estimates (RTX 4090 mini-PC 0.6–0.9 kW), which §II.E admits are not new measurements.
  • Renewable window T_d / synthetic trace renewable fraction = 2.5 h average (CAISO)
    The 7-day simulation trace is 'calibrated on CAISO curtailment statistics' but not released; the achievable 52% reduction is bounded by the trace's renewable fraction, and the 60% ceiling compared to Liu et al. [8] is inherited from that calibration.
  • γ, β utility weights = unspecified
    Defined in §VI.F for the utility model U(w,d) = γR(d) − βL(d) but never instantiated or swept in the evaluation, so the optimization stage of the orchestrator is not actually pinned down by the experiments.
assumptions (6)
  • domain assumption Single-GPU workloads with self-contained 24–40 GB checkpoints; distributed training states cannot be migrated as a unit
    Stated as a 'deliberate scope decision rather than a limitation' in §IV. All feasibility conclusions apply only to this workload class; the paper's own extension discussion (§VIII.D) concedes distributed workloads need different mechanisms.
  • domain assumption Stable effective WAN bandwidth between sites during migration
    Invoked throughout the feasibility filter and evaluation (§IV.C, §V.C, Algorithm 1). §VIII.F concedes 'background traffic and routing changes may temporarily reduce throughput' and only suggests 'conservative margins' as mitigation.
  • standard math Checkpoint transfer dominates migration cost: T_transfer = 8S/B with load ≈ 10.3 s, downtime ≈ 0.4 s
    The formula is definitional and the constants are drawn from cited systems ([17], [19], Table V); this is independent support, not circular.
  • domain assumption Curtailment windows are known to the scheduler at migration time (treated as deterministic 2.5 h in the evaluation)
    §VI.H introduces forecast uncertainty and a risk budget ε but the evaluation uses a fixed 2.5 h window; the orchestration benefit depends on the filter not overrunning the window.
  • domain assumption Hardware power figures and per-sample energies (Table I, §II.C) from vendor specs and unspecified 'prior work'
    §II.E explicitly states 'The system-level figures in this section are derived from publicly available specifications and consolidated measurements reported in prior work, rather than from new empirical measurements.' The ViT-B/32 2.7 mJ/sample figure and 'measurements from 2024–2025 deployments' are asserted without citable sources.
  • domain assumption The natural baseline for energy-driven scheduling ignores feasibility constraints (Energy-only policy)
    The Energy-only baseline is defined as migrating whenever renewables are available without checking transfer-time or window constraints (§VII.E), producing the +35% JCT penalty that motivates the proposal. This construction encodes the conclusion partly in the baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Green Distributed AI Training: Orchestrating Compute Across Renewable-Powered Micro Datacenters." pith.science (2026). https://pith.science/paper/KK4HL4EJ

@misc{pith2026251116182,
  author       = {Pith},
  title        = {Pith review of: Green Distributed AI Training: Orchestrating Compute Across Renewable-Powered Micro Datacenters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KK4HL4EJ}},
  note         = {Machine review of arXiv:2511.16182}
}
read the original abstract

The accelerating expansion of AI workloads is colliding with an energy landscape increasingly dominated by intermittent renewable generation. While vast quantities of zero-carbon energy are routinely curtailed, today's centralized datacenter architectures remain poorly matched to this reality in both energy proportionality and geographic flexibility. This work envisions a shift toward a distributed fabric of renewable-powered micro-datacenters that dynamically follow the availability of surplus green energy through live workload migration. At the core of this vision lies a formal feasibility-domain model that delineates when migratory AI computation is practically achievable. By explicitly linking checkpoint size, wide-area bandwidth, and renewable-window duration, the model reveals that migration is almost always energetically justified, and that time-not energy-is the dominant constraint shaping feasibility. This insight enables the design of a feasibility-aware orchestration framework that transforms migration from a best-effort heuristic into a principled control mechanism. Trace-driven evaluation shows that such orchestration can simultaneously reduce non-renewable energy use and improve performance stability, overcoming the tradeoffs of purely energy-driven strategies. Beyond the immediate feasibility analysis, the extended version explores the architectural horizon of renewable-aware AI infrastructures. It examines the role of emerging ultra-efficient GPU-enabled edge platforms, anticipates integration with grid-level control and demand-response ecosystems, and outlines paths toward supporting partially migratable and distributed workloads. The work positions feasibility-aware migration as a foundational building block for a future computing paradigm in which AI execution becomes fluid, geographically adaptive, and aligned with renewable energy availability.

Figures

Figures reproduced from arXiv: 2511.16182 by the authors.

Figure 1
Figure 1. Energy breakeven curves for checkpoint sizes from 1–100 GB. All breakeven points occur within minutes, confirming [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. We compare the Static, Energy-only, and Feasibility￾aware policies described in Section V. A. Feasibility Validation We first validate that the analytical feasibility boundaries of Section IV predict the migration overhead observed in the trace-driven simulator. For each workload class in [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 2
Figure 2. Feasibility domain: transfer-time isolines show that only sub-20 GB states migrate efficiently on 1–10 Gbps links. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 2 linked inside Pith

  1. [1]

    Turning curtailment into capacity: Why europe must rethink energy storage,

    Energy Central, “Turning curtailment into capacity: Why europe must rethink energy storage,” https://zenodo.org/records/15666971, 2024

    arXiv 2024

  2. [2]

    How europe’s grid operators are preparing for the en- ergy transition,

    Strategic Energy, “How europe’s grid operators are preparing for the en- ergy transition,” https://beyondfossilfuels.org/wp-content/uploads/2025/ 05/REPORT_FINAL.pdf, 2025

    2025

  3. [3]

    Nvidia jetson agx orin series,

    NVIDIA Corporation, “Nvidia jetson agx orin series,” NVIDIA, Tech. Rep., 2024, power envelope 15–60 W module; access date: 2025-11-20. [Online]. Available: https://www.nvidia.com/en-us/ autonomous-machines/embedded-systems/jetson-agx-orin/

    2024

  4. [4]

    Industrial edge-ai system based on nvidia jetson orin module,

    Edge AI System Vendor, “Industrial edge-ai system based on nvidia jetson orin module,” https://www.example-vendor.com/ jetson-orin-edge-ai, 2024, representative wall-plug power <150 W; access date: 2025-11-20

    2024

  5. [5]

    Introducing nvidia jetson agx thor — the ultimate platform for physical ai,

    NVIDIA Corporation, “Introducing nvidia jetson agx thor — the ultimate platform for physical ai,”NVIDIA Developer Blog, Aug. 2025, up to 2070 FP4 TFLOPS; 40–130 W envelope; access date: 2025-11-20. [Online]. Available: https://www.nvidia.com/en-us/ autonomous-machines/embedded-systems/jetson-thor/

    2025

  6. [6]

    Aetina jetson agx thor developer kit preview,

    Aetina Corporation, “Aetina jetson agx thor developer kit preview,” https://www.aetina.com/jetson-agx-thor-developer-kit, 2025, preview kit specification 100-150 W class; access date: 2025-11-20

    2025

  7. [7]

    Everfocus ear-100t robotics con- troller powered by nvidia jetson agx thor,

    EverFocus Electronics Corporation, “Everfocus ear-100t robotics con- troller powered by nvidia jetson agx thor,” https://www.everfocus.com/ ear-100t-jetson-thor, 2025, 130 W system budget robotics controller; access date: 2025-11-20

    2025

  8. [8]

    Renewable and cooling aware workload management for sustainable data centers,

    Z. Liu, M. Lin, A. Wierman, S. H. Low, and L. L. Andrew, “Renewable and cooling aware workload management for sustainable data centers,” inProceedings of the ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Systems, 2012, pp. 175–187

    2012

Show all 25 references
  1. [9]

    Free lunch: Exploiting renewable energy for comput- ing,

    s. Akoush et al., “Free lunch: Exploiting renewable energy for comput- ing,” inProceedings of the 13th USENIX Workshop on Hot Topics in Operating Systems (HotOS XIII). Napa, CA: USENIX Association, 2011

    2011

  2. [10]

    On the effectiveness of the ’follow-the-sun’ strategy in mitigating the carbon footprint of ai,

    R. Vergallo et al, “On the effectiveness of the ’follow-the-sun’ strategy in mitigating the carbon footprint of ai,”Preprint at Research Square, 2024, under Review

    2024

  3. [11]

    Energy optimal dispatch of the data center microgrid based on stochastic model predictive control,

    Y . Zhu et al., “Energy optimal dispatch of the data center microgrid based on stochastic model predictive control,”Frontiers in Energy Research, vol. 10, p. 863292, 2022

    2022

  4. [12]

    Towards sustainable ai: a comprehensive framework for green ai,

    A. Tabbakh et al., “Towards sustainable ai: a comprehensive framework for green ai,”Discover Sustainability, vol. 5, no. 408, 2024

    2024

  5. [13]

    Learning scheduling algorithms for data processing clusters,

    H. Maoet al., “Learning scheduling algorithms for data processing clusters,” inProceedings of ACM SIGCOMM, 2019, pp. 270–288

    2019

  6. [14]

    Follow the sun through the clouds: Application migration for geographically shifting workloads,

    Z. Shen et al., “Follow the sun through the clouds: Application migration for geographically shifting workloads,” inProceedings of the Seventh ACM Symposium on Cloud Computing (SoCC ’16). Santa Clara, CA, USA: Association for Computing Machinery, 2016, pp. 141–154

    2016

  7. [15]

    Speculative container scheduling for deep learning applications in a kubernetes cluster,

    Y . Mao et al., “Speculative container scheduling for deep learning applications in a kubernetes cluster,”IEEE Systems Journal, vol. 16, no. 3, pp. 3770–3781, 2022

    2022

  8. [16]

    Singularity: Planet-scale, preemptive and elastic scheduling of ai workloads,

    D. Shukla et al., “Singularity: Planet-scale, preemptive and elastic scheduling of ai workloads,” arXiv preprint arXiv:2202.07848, 2022, v2, submitted 21 Feb 2022. [Online]. Available: https: //arxiv.org/abs/2202.07848

    2022 arXiv

  9. [17]

    PhoenixOS: Concurrent OS-level GPU Checkpoint and Restore with Validated Speculation,

    X. Wei, “PhoenixOS: Concurrent OS-level GPU Checkpoint and Restore with Validated Speculation,” inProceedings of the 31st ACM Symposium on Operating Systems Principles (SOSP ’25). ACM, 2025. [Online]. Available: https://dl.acm.org/doi/10.1145/3731569.3764813

    2025

  10. [18]

    Llumnix: Dynamic scheduling for large language model serving,

    S. Sunet al., “Llumnix: Dynamic scheduling for large language model serving,” inProceedings of USENIX OSDI, 2024. [Online]. Available: https://www.usenix.org/conference/osdi24/presentation/llumnix

    2024

  11. [19]

    Serverlessllm: Fast and efficient serving for large language models,

    Y . Fuet al., “Serverlessllm: Fast and efficient serving for large language models,” inProceedings of USENIX OSDI, 2024. [Online]. Available: https://www.usenix.org/conference/osdi24/presentation/serverlessllm

    2024

  12. [20]

    ZeRO: Memory Optimizations Toward Training Trillion Parameter Models,

    S. Rajbhandari at al., “ZeRO: Memory Optimizations Toward Training Trillion Parameter Models,” inProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC ’20). ACM, 2020. [Online]. Available: https: //dl.acm.org/doi/10.11...

    2020 arXiv

  13. [21]

    Zero-offload: Democratizing billion-scale model training,

    J. Ren et al., “Zero-offload: Democratizing billion-scale model training,” inProceedings of the 2021 USENIX Annual Technical Conference (USENIX ATC). USENIX Association, 2021, pp. 551–564

    2021

  14. [22]

    A Study of Checkpointing in Large Scale Training of Deep Neural Networks,

    E. Rojas et al., “A Study of Checkpointing in Large Scale Training of Deep Neural Networks,” in2020 International Conference on High Performance Computing & Simulation (HPCS), 2020

    2020

  15. [23]

    Deep residual learning for image recognition,

    K. He et al., “Deep residual learning for image recognition,” inPro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  16. [24]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin et al., “Bert: Pre-training of deep bidirectional transformers for language understanding,” inProceedings of NAACL-HLT, 2019

    2019

  17. [25]

    Language models are unsupervised multitask learn- ers,

    A. Radford et al., “Language models are unsupervised multitask learn- ers,”OpenAI Technical Report, 2019. APPENDIXA ADDITIONAL MATERIAL A. Feasibility-Aware Migration Algorithm Algorithm 1 presents our feasibility-aware migration sched- uler. Algorithm 1Feasibility-Aware Migra...

    2019

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.