CENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives

C ENTILE: A Telemetry Foundation Model Evaluated by the Decisions It Drives Zifan Zhang* , Zhichao Hou* , Tingxiang Ji, Yuchen Liu

arXiv:2608.01725v1 [cs.NI] 3 Aug 2026

Department of Computer Science, North Carolina State University, Raleigh, NC 27695, USA

Abstract—Modern computing and networking infrastructure emits telemetry continuously, yet operators convert it into decisions with a separate predictor per task, entity, and horizon. One generative model, pretrained once over an operator’s own event streams, could replace this fleet, an approach that already scales to high-cardinality streams in recommendation systems. However, point-forecast error on operational telemetry saturates near simple last-value baselines, so lower error alone need not improve the decisions it feeds. To close this gap, we present C ENTILE, a generative foundation model for network and systems telemetry, evaluated by replaying the decisions its calibrated conditional quantiles drive. C ENTILE treats heterogeneous telemetry as event-driven, irregularly timed entity streams and serves flexible forecast horizons in a single pass, requiring no future timestamps. To our knowledge, C ENTILE is the first pretrained telemetry model to improve both HPC scheduling and network provisioning decisions under replay, its runtime estimator transferring zero-shot across months and its pretrained weights across domains from hours of target data. Extensive experiments on HPC job logs and network traffic confirm that C ENTILE lowers the mean bounded slowdown of backfilling by up to approximately 77% over deployed user estimates and roughly halves the deployed rule’s violation rate. Our code is available at https://github.com/ZzZTripleZzZ/all-in-one.

I. I NTRODUCTION Every consequential action in a production fleet begins as a forecast. An HPC scheduler reserves thousands of nodes against a predicted job runtime, and a capacity planner locks in next-hour bandwidth at a predicted traffic percentile. When the forecast is wrong, the cost is concrete, as idle reservations waste node-hours, under-provisioned links violate service agreements, and tail latency climbs at fleet scale. The data for these forecasts is abundant, since every layer of modern computing and networking infrastructure emits telemetry continuously, from per-task lifecycle events and pervirtual-machine utilization in datacenters, to per-node power in supercomputers, to per-flow traffic volumes in networks spanning datacenter fabrics and national backbones [1], [2], [3], [4]. These events arrive irregularly, on each entity’s own clock rather than on a shared sampling grid. The prevailing practice converts this telemetry into decisions one predictor at a time, with a separate model trained for each task, each entity type, and even each prediction horizon. This is precisely the regime that foundation models displaced in language, vision, and recommendation, yet the telemetry prediction that underpins network performance analysis still * These authors contributed equally to this work.

lacks one. We therefore ask whether systems telemetry can have its own foundation model, a single pretrained model that ingests any entity’s history and serves any operational decision that consumes it. Three requirements make this question concrete: the model must 1) serve any forecast horizon without retraining; 2) consume irregularly timed events as they arrive; and 3) carry what it has learned across domains, so that telemetry from one infrastructure can bootstrap decisions on another. One pretrained model would retire this fleet, and a newly deployed system with only hours of history could still make calibrated decisions on day one. Building the first such model is the goal of this paper. Motivation and Challenges: Reaching this goal is difficult for reasons specific to networking systems. First, the fleet itself is a recurring operational cost. Each predictor carries its own training pipeline, retraining cadence, storage and inference footprint, and monitoring, an operational burden documented to dominate the lifecycle cost of production learning systems [5], and every new workload, entity type, or horizon adds another model and pipeline. Second, the data itself is difficult to capture with a single model, as telemetry entities number in the millions and are heterogeneous, spanning users, jobs, machines, and flows. Moreover, events do not arrive on a uniform clock, a substantial part of their predictive signal is carried by when they happen, and operators query these streams at arbitrary input and output lengths, so a model that fixes its window and horizon must be duplicated per use. Lastly, and most fundamentally, today’s models are judged by a metric that is insensitive to what matters. Actions speak louder than words: a telemetry model exists to drive schedules and provisioning plans, yet it is ranked by point-forecast error, a number that saturates near last-value baselines on operational traces and can stay flat while the decisions it feeds diverge [6]. Limitations and Opportunities: The foundation models proposed for networking so far each cover one domain, one modality, and usually one task. Masked packet and flow encoders reconstruct held-out telemetry yet are never evaluated against a scheduler or a provisioner [7], [8], wireless models remain confined to their own air interface [9], [10], and language-model adaptations inherit billion-parameter inference costs that no per-decision millisecond budget can absorb [11]. General-purpose time-series foundation models [12], [13] come closest, delivering one model for any horizon, but they pretrain on generic corpora rather than system telemetry,

Telemetry entities

Replayed decisions

Centile

HPC jobs ¢ per user

events 1

ISP traffic ¢ per link

Event adapter

EASY backfilling

Student-t mixture head Centile attention

value ¢ covariates

2

3

causal £L

¢t ¢ clock

quantiles

w^ = clip(Q q )

Provisioning

any q, any h

DC tasks ¢ per VM

capacity = Q ¿

5

Stage I: pretrain once

Stage II: serve in one pass

6

irregular gaps ¢t carry signal

One attention layer

Z

pointwise attention

split

Á2 (QK > +B) V

f1 ; Á1

§IV-D

zero-shot transfer

£L Norm

2

§IV-B

4

no softmax

¯

output

©

f2

Y

update gate

position/time bias

suppresses stale history

Decision-serving data flow 1

events: per-entity (ti ; xi ) + submit covariates ci + 1

2

3

replay: w^ = clip(Q q ) backfill ¢ capacity Q ¿

summarize: £L attention layers ! state zi

decode: Student-t mixture ! any Q q , one pass

4

5

pretrain once: next-event likelihood objective

6

zero-shot: frozen weights serve later months (§IV-B)

Fig. 1. C ENTILE overview. Events from telemetry entities flow through the adapter, the pretrained attention layers, and the Student-t mixture head into calibrated quantiles that drive the replayed decisions. The inset expands one of the L attention layers.

job logs, and provisioning next-hour capacity on national ISP traffic. A third domain, cloud virtual-machine telemetry, joins the pretraining traces for cross-domain transfer. To our knowledge, this is the first work to show that one model design, pretrained on unlabeled telemetry, improves both scheduling and provisioning decisions under replay. The estimator transfers zero-shot across months, and the pretrained weights carry over from traffic prediction to HPC scheduling. Our main contributions are summarized as follows.

assume a uniform sampling grid that discards event timing, and serve fixed quantile grids whose most conservative level falls short of what provisioning requires. None is at once generative, pretrained on system telemetry itself, and evaluated on the decisions operators actually run. Our key insight is that operational decisions consume quantiles rather than point forecasts, as a backfilling scheduler reserves with a walltime estimate and a provisioner sets capacity at a percentile. A generative model that learns the calibrated conditional distribution of telemetry is therefore exactly the artifact both decisions need, even where its point error appears unremarkable. Objectives and Contributions: To this end, we present C ENTILE1 , a generative foundation model for network and systems telemetry that is evaluated throughout at the level of the operational decisions it drives. C ENTILE treats heterogeneous telemetry as event-driven entity streams and embeds each event’s value, resource covariates, and inter-event gap. From one shared design, the same architecture, objective, and training procedure at every deployment, it pretrains one model per domain, built around three telemetry-specific choices: intensity-preserving attention over irregularly timed event streams, a heavy-tailed mixture output, and direct multihorizon decoding. C ENTILE exposes calibrated conditional quantiles that feed each decision, serving every horizon in a single pass rather than by autoregressive rollout and needing no future timestamps at submit time, as sketched in Fig. 1. We instantiate this on two operational tasks that no prior telemetry model has jointly addressed: estimating job walltimes, the runtimes that Extensible Argonne Scheduling sYstem (EASY) backfilling schedules against, on production HPC

We recast network and systems telemetry prediction as one generative model over event-driven entity streams, and introduce a decision-replay protocol that scores a telemetry model by the scheduling and provisioning decisions its forecasts drive rather than by point-forecast error. This protocol exposes differences that point-forecast metrics conceal. • Driving EASY backfilling from C ENTILE quantiles lowers mean bounded slowdown over deployed user estimates by up to approximately 77%, and its estimator is the best walltime source on every evaluated month, including the unseen months served by the April-pretrained model. • Without any change to the design, C ENTILE serves nexthour capacity provisioning on ISP traffic, roughly halving the violation rate of the deployed rule at comparable overprovisioning. Its 1.3M-parameter runtime estimator serves scheduling decisions in milliseconds on commodity CPUs. • Lastly, we conduct comprehensive decision-replay evaluations across a multi-month HPC job trace and a national ISP trace, spanning cross-month transfer, cross-domain transfer, pretraining scale, ablations, and deployment cost. •

1 Our code is available at https://github.com/ZzZTripleZzZ/all-in-one.

2

reservation, or if it fits entirely on nodes the reservation leaves spare. The estimates therefore govern both the reservation and every backfill decision. We score a schedule by bounded slowdown [19],   wait + run ,1 , (2) BS = max max(run, ϵ)

Table I. Positioning against foundation models for systems and time series. ✓: yes, ✗: no, ∼: partial. Model

Domain

Gen.

Sys-pretr.

ET-BERT [7], netFound [8] packets NetLLM [11] networking MobiGPT [9], UoMo [14] wireless LWM [10] wireless Chronos [12], TimesFM [13] generic TS C ENTILE (ours) HPC+ISP

✗ ✓ ∼ ✗ ✓ ✓

✓ ✗ ✓ ✓ ✗ ✓

Any-hor. Multi-task ✗ ✗ ∼ ✗ ✓ ✓

✗ ✓ ✓ ✗ ✗ ✓

Time-aware ✗ ✗ ✗ ✗ ✗ ✓

where wait and run are a job’s queue wait and running time, and ϵ is a short floor that keeps very short jobs from dominating the ratio, with lower values preferred. Capacity provisioning consumes a forecast differently, as the capacity for the next interval is set to the served quantile Qτ of predicted demand, and the operator moves τ to trade overprovisioning against the rate at which demand exceeds the provisioned capacity. In both cases, we evaluate by decision replay, re-executing the recorded decision process on the historical trace with the served estimate substituted for the deployed one, holding arrivals, true runtimes, and job sizes fixed, and scoring the decisions that result.

II. BACKGROUND AND P OSITIONING Telemetry as entity event streams: We model all telemetry through one abstraction. A source exposes a set of entities E, where an entity may be a user, a compute node, a virtual machine, or a monitored link. Each entity e ∈ E emits an ordered stream of timestamped events
Se = (t1 , x1 ), (t2 , x2 ), . . . , t1 < t2 < · · · , (1) where xi records the measurement observed at time ti , such as a demand volume, a utilization sample, or an event duration, together with any covariates the source attaches. The timestamps are the entity’s own, so the inter-event gaps ∆i = ti − ti−1 vary over orders of magnitude rather than following a fixed sampling period. Over such streams, operators pose two recurring questions: forecasting the distribution of the next measurements at one or many horizons and, downstream of that forecast, the operational decisions that consume its quantiles, from scheduling batch jobs to provisioning capacity. Positioning: Table I places C ENTILE against representative foundation models for systems and time series. Its columns ask whether a model is generative in the sense of exposing a full predictive distribution, is pretrained on system telemetry, serves arbitrary output horizons, reuses one model design across tasks, and models irregular inter-event gaps. No prior model is at once generative, pretrained on system telemetry, horizon-agnostic, reusable across tasks, and time-aware, and C ENTILE is the first to satisfy all five. Key observation: telemetry is bursty in time yet correlated in value: We measure, per entity, the coefficient of variation (CV) of inter-event gaps and the lag-one autocorrelation of the target across five public traces (Table II). Inter-event timing varies widely, as job and task logs [15], [3], [16] reach a CV between one and four and node sampling punctuated by downtime [2] reaches nearly twelve, while only the fixedrate Azure sampler [1] attains the zero of a strict grid. At the same time, targets are frequently autocorrelated, up to 0.93. Both signals, the value history and the irregular timing, are available to a model that reads the stream in order yet discarded by a uniform-grid forecaster, which motivates a time-aware generative model of telemetry. From forecasts to decisions: We specify how each replayed decision consumes a served quantile. Under EASY backfilling [17], jobs queue in arrival order, and the scheduler makes a reservation for the queue head by projecting when running jobs will release their nodes from the walltime estimates those jobs carry [18]. A waiting job may jump ahead of the queue head only if its own estimate indicates it will finish before the

III. T HE C ENTILE F RAMEWORK C ENTILE is a generative foundation model over per-entity telemetry event streams, assembled from three composable stages: a pluggable input adapter that embeds each event, a shared set of attention layers that summarize an entity’s history, and a distributional head that emits the next measurement as a full predictive distribution. We design C ENTILE to learn the conditional law of the next event an entity emits, p(xi+1 | x≤i ), so that it can roll a stream forward to whatever horizon an operator queries. Prior networking models instead use masked traffic encoders [7], [8] or frozen language models adapted to networking tasks [11], neither of which exposes such a distribution. Design goals: Four goals shape the framework: 1) one artifact per domain, meaning any entity history and any horizon within a domain are served by a single set of weights, and a new domain reuses the design unchanged; 2) calibrated distributions, meaning the served quantiles carry their nominal coverage so that a decision policy can choose its operating point a priori; 3) submit-time causality, meaning every prediction is computable from information available when the decision fires, with no future values or timestamps; and 4) negligible inference cost, meaning milliseconds per decision on the commodity CPUs that already run the scheduler and the provisioning loop. Together, these goals operationalize the introduction’s three requirements while adding the calibration and inference-cost constraints that an operational control plane imposes. Table II. Telemetry is irregular yet predictable. Median per-entity inter-event CV and lag-one autocorrelation. Larger CV means less clock-like timing.

3

Dataset

Type

F-DATA (HPC jobs) [15] Borg (DC tasks) [3] Alibaba (DC tasks) [16] Azure (VM CPU) [1] M100 (node power) [2]

job log task log task log sampling sampling

Inter-event CV

Autocorr.

3.97 3.08 1.03 0.00 11.94

0.48 0.82 −0.06 0.09 0.93

A High-Level Walkthrough: Figure 1 sketches the pipeline, and its numbered steps trace one decision end to end. The input adapter maps each incoming event to a D-dimensional embedding of its value and the covariates of its deployment mode (step 1). The L attention layers aggregate the causal history and return a state that summarizes the entity up to the current event (step 2). The distributional head turns this state into a Student-t mixture, from which any quantile is estimated empirically from Ns Monte Carlo draws, as the mixture admits no closed-form inverse (step 3). The served quantile becomes the walltime estimate or the provisioned capacity of the replayed decision (step 4). In forecaster mode, a horizon feature conditions the head, so one pass yields a predictive distribution for every horizon at once. In estimator mode, the model instead reads the final position of the current job’s history and serves one clipped quantile as the walltime estimate. The weights are pretrained once (step 5) and reused on later months zero-shot (step 6).

B. Model Architecture Input adapter: Each event arrives as a feature vector that the adapter embeds linearly into width D. In estimator mode, the vector carries the entity’s previous observed runtime, lagged and log-compressed, together with the job’s submit-time request, its node count and declared limit, and an extended configuration adds the logarithm of the inter-arrival gap and clock features, the sine and cosine of the hour of day and the normalized day of week. In forecaster mode, the adapter pairs a projection of each observed value with a learned embedding of its quantile bin, one of Kc (e.g., 256) bins spaced by the training-region quantiles, which adds a discrete regime signal to the continuous magnitude. Targets are standardized in log space, keeping the heavy-tailed job runtimes and link traffic volumes on a scale the head models well. Because the adapter is the only source-specific stage, a new telemetry source, whether a switch counter feed or a per-virtual-machine utilization stream, is added by declaring its feature vector while the stages above it stay fixed. Intensity-preserving attention: The L attention layers are built around a telemetry-specific principle, namely that the absolute intensity of activity is itself informative. A standard Transformer’s softmax attention [21] renormalizes every history to sum to one, which flattens a submission burst and an idle period into the same relative weights. Each layer therefore scores the causal past pointwise [22] and lets the aggregate scale with the entity’s level of activity. Replacing the softmax mixer is an established route to new attention properties, such as robustness [23], and here it targets intensity preservation. From the layer input Z, four projections produce a gating branch, a value branch, and the query–key pair,

A. Problem Formulation A telemetry source exposes a set of entities, and each entity emits an ordered stream Se of timestamped events as in (1). Writing Hi = (x≤i , t≤i+1 , ci+1 ) for what the control plane holds when the next event is submitted, C ENTILE models the one-step predictive distribution p(xi+1 | Hi ) ,

(3)

where x≤i is the observed value history, t≤i+1 collects the timestamps through the submit time of the next event, and ci+1 holds the submit-time covariates the scheduler already knows before the event completes, such as the requested node count and the user-declared limit. In forecaster mode, the same construction is conditioned additionally on a requested horizon h, which §III-B realizes without rollout. Writing F̂i for the cumulative distribution function of (3) and Qq (F̂i ) for its qquantile, the model is calibrated at level q when the served quantile carries its nominal coverage, h i Pr xi+1 ≤ Qq (F̂i ) Hi = q, (4)

U (Z), V (Z), Q(Z), K(Z) = Split(ϕ1 (f1 (Z))) ,

(6)

where f1 is a linear map and ϕ1 is the sigmoid linear unit (SiLU) nonlinearity. The layer then weights the causal past through pointwise SiLU attention scores, forgoing softmax normalization, and gates the pooled values elementwise,


Y = f2 Norm ϕ2 QK ⊤ + B V ⊙ U , (7) where ϕ2 is again SiLU, ⊙ denotes the elementwise product, Norm is layer normalization, and f2 is the output projection. The learned bias B carries a relative position term that weights events by their order in the stream and an optional time term, multi-scale Fourier features of the causal gap log(1 + ti − tj ) projected per head, which forecaster mode enables throughout so that the layer reads irregular spacing directly. The four branches divide the computation, where Q and K select which past events are relevant to the one being processed, V carries what each past event contributes to the summary, and U decides how much of the pooled history the current event admits. On a job stream, for instance, a submission burst raises the pairwise scores among its own events and the aggregate grows with the burst, while after a long idle period the gate admits little of the now-stale history. A causal mask confines every position to its own past, and LayerScale [24] stabilizes

a property whose aggregate consequence we audit on heldout data by comparing nominal levels with empirical coverage [20]. Calibration is the property both decisions consume, rather than point accuracy. For walltime estimation, the predictive distribution is reduced to one served value by a quantileand-clip rule,  
ŵ = clip Qq F̂i , wmin , wuser , (5) where wmin is a short minimum-runtime floor, for instance 60 seconds, and wuser is the user-declared limit, which also remains the kill limit at which an overrunning job is terminated. For capacity provisioning, the reduction is simpler, as the capacity provisioned for the next interval on a link is the served quantile Qτ (F̂i ) of predicted demand.

4

where B indexes the scored positions in a batch, θ collects the adapter, attention, and head parameters, and pθ is the mixture of (8) evaluated at the realized measurement xi+1 . The log-likelihood is a strictly proper scoring rule [20], so its minimizer over a sufficiently rich family is the true conditional distribution rather than any single point summary, which is what a decision consuming a quantile needs. After training, the weights and the log-space normalization statistics are frozen together, so a later deployment window enters the same scale the head was fit on. At inference, the estimator fetches the entity’s most recent C events, appends the submit-time covariates ci+1 , and runs one pass through the model. It then estimates Qq (F̂i ) by Monte Carlo sampling from the mixture and returns the clipped estimate of Eq. (5) within a single scheduling cycle. Training reads only data strictly earlier than the evaluation window, and each entity’s history grows in submit order at inference, so no prediction ever reads its own future, a contract every baseline in §IV obeys as well.

optimization at depth. One unmodified layer stack thus serves both deployment modes, and the defaults in §IV state which timing inputs each replay serves. Inference cost scales as the attention product over a context of bounded length, so the per-decision compute is a small constant, as the measured milliseconds of §IV-F confirm, and the estimator runs on the CPU-only control plane that hosts the scheduler. Distributional head: The head parameterizes a mixture of M Student-t components [25], an expressive form for the heavy right tails that job runtimes and traffic volumes exhibit. From the state zi , the head emits mixture weights αm , locations µm , scales σm , and degrees of freedom νm for m=1, . . . , M , and assembles the predictive density p(xi+1 | zi ) =

M X

αm T (xi+1 ; µm , σm , νm ) ,

(8)

m=1

where a softmax makes the weights αm sum to one, a softplus keeps each scale σm > 0, and a positive offset keeps each νm > 2 so the component variance stays finite. Each component is the location-scale Student-t density
"  2 #− ν+1 2 Γ ν+1 1 x − µ 2
1+ , (9) T (x; µ, σ, ν) = ν √ ν σ Γ 2 πν σ

Pretraining and zero-shot transfer: Because the estimator is pretrained once rather than fit per deployment, a single artifact can be reused across time. We pretrain on earlier months and deploy the frozen artifact on later months, which supply only their own per-entity histories as context, so successive months are served as they arrive without retraining.

where Γ is the gamma function and a smaller ν yields a heavier tail than a Gaussian of the same scale, letting the head place mass on long runtimes and rare traffic spikes. Unlike a quantile-regression head [26], which fixes a grid of levels at training time and must be refit when the decision moves its operating point, the mixture head represents the full distribution once and serves any quantile on demand. Direct multi-horizon conditioning: In forecaster mode, the horizon h enters as Fourier features of log(1 + h) that are projected and added to the state, so a single pass produces the distribution at every horizon. Writing zi for the state at event i, the head reads a horizon-shifted state z̃i (h) = zi + Wψ ψ(log(1 + h)) ,

D. Why Calibrated Quantiles Improve Decisions In this subsection, we provide a theoretical account of the design by establishing how the calibration property of Eq. (4) converts into guarantees on the two replayed decisions and why direct decoding avoids the compounding penalty of rollout at long horizons. The analysis rests on two assumptions that are standard in probabilistic forecasting and in the analysis of learned dynamical models [20], [27], [28], and both are attainable in deployment rather than idealizations. Calibration is what the objective of Eq. (11) already pursues, since a strictly proper scoring rule is minimized by the conditional law itself, and the served level is audited against measured coverage before deployment rather than assumed. The rollout comparison needs only a bounded expected one-step error and a Lipschitz feedback map.

(10)

where ψ maps the scalar log(1 + h) to sinusoidal Fourier features and Wψ projects them to width D before the head decodes z̃i (h). Because the horizon enters only through this additive term, the layers run once and each requested horizon costs one more projection rather than one more rollout step, and §III-D quantifies the gap to autoregressive rollout.

Assumption 1 (Quantile calibration). At the served level and horizon, the served distribution satisfies Eq. (4) given the prediction-time history, for the events the decision consumes. Under Assumption 1 at level τ , demand exceeds the capacity provisioned at the served quantile at the promised rate 1 − τ . The classical newsvendor argument then identifies the costoptimal level as the fractile κu /(κu + κo ) of the per-unit violation cost κu and overprovisioning cost κo , since the expected cost is convex in the provisioned capacity and its derivative changes sign at that fractile [29]. Calibration at the served level is what carries this first-order condition from the unknown demand law over to the distribution C ENTILE serves. Thus, the served quantile becomes a contract, and §IV-D tests it empirically, where measured violation rates stay at or below

C. Training and Inference C ENTILE is pretrained once in an offline pipeline and then queried once per decision, which keeps the data-intensive stage off the control loop the scheduler and the provisioner run. Pretraining groups the archived telemetry into per-entity streams in submission order and scores every event against the history that preceded it, so each position predicts its entity’s next measurement from its own causal past by minimizing 1 X L(θ) = − log pθ (xi+1 | zi ) , (11) |B| i∈B

5

the promised 1 − τ at every served level while a miscalibrated baseline exceeds it at two of three served levels.

Remark. The guarantees are conditional on measured calibration rather than unconditional properties of the architecture, and the rollout comparison rests on the stated regularity of the feedback map. Once measured, calibration makes the guarantees above operative up to the Monte Carlo error of quantile estimation.

Proposition 1 (Reservation reliability under EASY). Consider the J jobs backing a reservation, each with a walltime estimate served by Eq. (5) at level q upon submission. Under Assumption 1, the probability, assessed at submission, that any of them overruns its served estimate is at most J(1 − q), and the clipping in Eq. (5) never degrades this coverage.

IV. E XPERIMENTS A. Experimental Setup We evaluate C ENTILE at the decision level on two tasks: 1) walltime estimation for EASY backfilling on F-DATA [15], production job logs from the Fugaku supercomputer; and 2) capacity provisioning on CESNET-TimeSeries24 [30], hourly traffic volumes from a national ISP backbone network. The primary F-DATA month contains roughly 117K jobs, of which the replay window covers 48,826 jobs from 298 users over two weeks, following a one-week warm-up that fits the estimator and seeds per-user histories. The median user-declared limit, six hours, exceeds the median true runtime, roughly 21 minutes, by more than an order of magnitude. CESNET is split chronologically at 80/20 per series and evaluated on 3,000 held-out windows.

Proof sketch. For one job, if wmin ≤ Qq ≤ wuser the served estimate equals Qq and the overrun probability is 1 − q by Eq. (4). Otherwise the estimate is wuser , which the runtime cannot exceed because the kill limit enforces it, so the overrun probability is zero. Raising the estimate to the floor wmin only enlarges coverage. A union bound over the J backing jobs gives the claim. In other words, with a calibrated q, the projected noderelease times behind a reservation are simultaneously conservative with probability at least 1 − J(1 − q), a guarantee that becomes operative at conservative served levels where J(1−q) stays small. Read as a sizing rule, an operator targeting reservation reliability 1−δ serves at q ≥ 1−δ/J. At the same time, the estimates stay well below the deployed limits and keep backfill holes open. The architecture of §III-B enters here, since the bound is only as useful as the distribution behind it. The mixture head keeps a heavy right tail per job rather than one global margin, and the intensity-preserving attention supplies the history that separates a user’s long jobs from the short ones. Thus, the conditional quantile is conservative precisely for the jobs whose histories predict long runtimes and aggressive elsewhere, a per-job adaptivity no single point estimate provides, and the replay of §IV-B measures its value at moderate served levels. Capacity provisioning consumes the same guarantee in a different currency, where 1 − τ is the fraction of intervals in which a link carries more than the capacity provisioned for it, so the served level states a service-level target rather than a tuned constant. Direct decoding carries a guarantee of its own, under the standard regularity for autoregressive prediction, namely a bounded expected one-step error and a Lipschitz feedback map [27], [28]. Rollout feeds each prediction back as input, so the error at every step decomposes into a fresh one-step term plus the amplified error it inherited. Unrolling this recursion bounds the expected rollout error at horizon h by a geometric sum in the Lipschitz constant of the feedback map, a bound that stays finite only under a strict contraction, grows linearly in h when the constant is one, and grows exponentially beyond it. Bursty telemetry offers no certificate of contraction [27]. The direct head of Eq. (10) removes the compounding term entirely, as each horizon is a separate target read from the same observed history, so no prediction is ever fed back and each horizon incurs only its own estimation error. Thus, the widening gap measured in §IV-E is consistent with the compounding the bound predicts rather than a tuning effect.

Decision replay: Each F-DATA month is replayed through an EASY backfilling scheduler in which the walltime estimate is the only quantity that varies between the compared estimators, and the kill limit remains the user-declared value. Reservations apply the standard estimate-extension correction [18], extending an expired estimate in place rather than killing the job. Cluster capacity is set from the offered load of the replayed window, and we evaluate the backlogged band (offered-tocapacity ratios between 0.95 and 1.46), where queues form. We report the mean and the 95th percentile of bounded slowdown as defined in §II. Forecast quality is reported through per-series Mean Absolute Scaled Error (MASE) [31], for which 1.0 matches the naive last-value forecast, pinball loss at the 0.99 quantile [26], and a nine-quantile Continuous Ranked Probability Score (CRPS) approximation [20]. Estimator baselines: We compare five walltime sources: 1) the user-declared limit, the operator status quo; 2) the peruser rolling 90th-percentile of past runtimes (History q90); 3) the average of the user’s last two runtimes (Tsafrir last2), the canonical system-generated predictor [18]; 4) gradientboosted quantile regression on causal tabular features (GBM quantile), the strongest non-sequential learned baseline in recent scheduling studies [32], [33]; and 5) C ENTILE, serving the requested quantile of its predicted runtime distribution at submit time. Scheduling with true runtimes is a reference, and all walltime sources observe identical causal histories. Implementation and defaults: Models are implemented in PyTorch and trained with bf16 autocast on one A30 GPU, and per-user histories are maintained strictly in submission order during both training and replay. The runtime estimator uses width D=256, L=4 attention layers, eight attention heads, a context of C=96 events, M =6 mixture components, Ns =2048 Monte Carlo draws per served quantile, and 30

6

Mean bounded slowdown

Table III. Bounded slowdown by walltime source at two backlogged capacities on the primary month. Bold marks the best estimator, and true-runtime scheduling is included only as a reference.

Walltime source User estimates History q90 Tsafrir last-2 GBM quantile C ENTILE True runtimes (ref.)

20K nodes

23K nodes

mean

mean

P95

53.1 37.6 23.1 19.7 15.5±1.1

171.4 53.9 21.2 32.1 17.8

28.9

169.9

P95

94.8 279.9 56.6 104.8 24.3 31.1 25.4 34.9 22.2±2.1 28.0 17.7

69.6

100

50 30 20 18K

20K

21K

22K

23K

Cluster capacity (nodes)

Fig. 2. Mean bounded slowdown across the backlogged capacity band on the primary month. C ENTILE improves on the deployed estimates and the history baseline at every capacity.

training epochs, serving q=0.5 in the primary tables and q=0.6 in the capacity and month sweeps, and the CESNET forecaster uses D=512 and L=8. C ENTILE rows in Table III report the mean and standard deviation over five seeds. The primary and month replays serve the estimator in its minimal configuration, which reads the previous runtime and the request covariates without event-timing inputs, as richer timing inputs did not improve this replay. The transfer and pretraining-scale protocols serve the extended covariates and the attention time term, and the arms of §IV-E are covariate-matched as noted there.

C ENTILE improves on the deployed estimates and the history baseline at every capacity and surpasses the true-runtime reference at the 23K edge (16.1 versus 28.9). One pretrained estimator transfers across months: Table IV and Fig. 3 extend the replay to five further months at representative backlogged operating points. We observe that C ENTILE is the best estimator in every column. The July and October columns are served by the zero-shot model, pretrained on April and never shown the target month, while the remaining columns fit the estimator on their own warm-up windows. This advantage can be attributed to the pretraining data, which is larger and more diverse than any single month’s warm-up window. The advantage holds at the P95 tail in every column, adding no tail risk.

B. Backfilling Replay on F-DATA C ENTILE schedules near, and beyond, the true-runtime reference: Table III reports the primary month at two capacities in the backlogged band. It is observed that C ENTILE attains the lowest mean bounded slowdown among all estimators at both capacities, 22.2 and 15.5 versus 24.3 and 23.1 for the strongest history-based estimator, respectively, improving on the deployed user estimates by up to approximately 77%. At the 23K-node capacity C ENTILE also surpasses scheduling with true runtimes (15.5 versus 28.9), and its P95 tail is the best of every walltime source, reference included, at both capacities. This is because schedule quality under EASY is not monotone in estimate accuracy, a property established well before learned estimators [18]. Estimates far above the true runtime, as the deployed user limits are, block backfilling outright because an inflated request no longer fits the holes a reservation leaves. Exact runtimes sit at the opposite extreme, where reservations receive no protective slack and the scheduler forfeits the heel-and-toe margin by which slightly conservative estimates keep short jobs flowing into shrinking holes [18]. The optimum therefore sits between the two, at estimates that are accurate yet slightly conservative for the jobs that back reservations, which is precisely the operating point a calibrated conditional quantile provides. Consequently, the true-runtime reference does not bound the schedule quality attainable under EASY, and it is not monotone in cluster size either, since the backfill opportunities that carry the effect grow with the free capacity. Its tail makes the same point, as the reference P95 rises from 69.6 to 169.9 across the two capacities while every estimator improves, the signature of a few queue-head jobs repeatedly displaced rather than of a uniformly worse schedule. Figure 2 extends the replay to five capacities across the backlogged band at q=0.6, where

Impact of pretraining scale: Enlarging the pretraining data from roughly 148K to 1.07M jobs lowers the mean bounded slowdown at the moderate-overload point by approximately 35% (11.9 to 7.7). The largest pretraining set of roughly 1.8M Table IV. Mean bounded slowdown across F-DATA months at representative backlogged operating points. † marks months served zero-shot by the April-pretrained estimator. Walltime source

Jun

Jul† Aug

Sep

Oct†

User estimates History q90 Tsafrir last-2 GBM quantile C ENTILE

3.68 2.54 3.32 3.35 1.86

66.3 58.2 58.1 54.4 42.6

21.9 14.2 10.5 11.7 7.59

14.0 6.09 5.85 4.15 3.78

Mean bounded slowdown

User estimates History q90

31.7 10.1 9.08 14.6 7.35

Tsafrir last-2 GBM quantile

Centile

50 20 10 5 2 Jun

Jul

Aug

Sep

Oct

Fig. 3. C ENTILE attains the lowest mean bounded slowdown on every month. Shaded months are served zero-shot by the April-pretrained model.

7

Table V. Cross-domain transfer. Mean bounded slowdown at the 20,000-node capacity (three-seed mean±std) as the target-data budget shrinks, with Transfer initialized from network and cloud telemetry alone. Bold marks the better arm. Target budget

Scratch

Transfer

6 hours 1 day 2 days 7 days (full)

83.0 ± 36.0 43.7 ± 13.1 25.8 ± 1.2 26.0 ± 2.2

35.9 ± 1.3 22.8 ± 2.3 24.6 ± 2.9 26.0 ± 1.5

Served quantile τ

0.5

0.8

0.9

0.95

Chronos-Bolt zero-shot [12] C ENTILE

48.7 42.8

23.0 17.5

13.1 9.0

— 4.7

50

38.2

Violation rate (%)

Statistics only, no training

Table VI. Provisioning violation rate (%) at matched served quantiles on CESNET. C ENTILE incurs fewer violations at every τ and serves quantile levels beyond the zero-shot model’s grid.

jobs attains the best light-load slowdown (2.4, roughly 25% below the smallest set). C. Cross-Domain Transfer

40 30 20 10 0

Setup: The zero-shot months above hold the domain fixed, so we now ask whether the pretrained weights transfer across domains. We pretrain the estimator design in its extended configuration on source traces of per-institution hourly traffic and per-virtual-machine CPU telemetry, neither of which contains a single job duration, cast into the estimator’s event format with per-domain standardization. These weights then initialize the F-DATA walltime estimator, fine-tuned on the last six hours, one day, two days, or all seven days of target history, then replayed at the 20,000-node capacity as in §IV-B. Three arms share every remaining setting: a scratch estimator trained from random initialization at each budget, a transfer estimator initialized from the cross-domain checkpoint, and a statistics-only arm that loads the checkpoint and fits only normalization statistics. The history each arm may read at inference is identical, every entry is a three-seed mean with standard deviation, and the budget protocol yields numbers slightly above the five-seed results of Table III. Cross-domain pretraining makes scarce-data training reliable: We observe that at the six-hour data budget the three scratch seeds are highly inconsistent, with mean bounded slowdown ranging from 36.7 to 124.5, so a single from-scratch training run at this budget is not auditable in advance. The transfer arm is instead consistent across seeds, 35.9 ± 1.3 (Table V), and lowers the served-quantile pinball loss by roughly 35% at every budget below the full window. This reliability can be attributed to the cross-domain checkpoint, which supplies an event-stream prior that random initialization cannot recover from a few hundred jobs, and a checkpoint pretrained on the network stream alone performs comparably. As the budget grows the two arms converge, 24.6 versus 25.8 at two days and parity at seven days. One day of target data suffices: Notably, the transfer estimator fine-tuned on a single day of target history reaches 22.8 versus 43.7 for scratch, and the worst transfer seed (26.1) outperforms the best scratch seed (33.9). One day of transfer already lands below the protocol’s own seven-day arms and well below the deployed estimates (94.8) and the history baseline (56.6). The statistics-only arm yields 38.2, below both deployed references, because the transferred weights already encode the heavy-tailed event structure the decision consumes.

20

40

60

80

100

Overprovisioning (%)

Fig. 4. Provisioning frontier on CESNET. Crosses denote operator rules, and the zero-shot model’s grid ends at τ =0.9.

D. Capacity Provisioning on CESNET Calibrated quantiles halve provisioning violations: Figure 4 replays next-hour capacity provisioning, sweeping the served quantile τ against the rules operators deploy in practice. Notably, at an overprovisioning level comparable to the Exponentially Weighted Moving Average (EWMA) rule, C ENTILE cuts the violation rate from 20.5% to 9.0%, and every C ENTILE operating point improves on the operator-rule frontier. Table VI shows that C ENTILE yields fewer violations than a zero-shot time-series foundation model at every τ the baseline serves, and only the calibrated model keeps its violation rate within the promised budget. Furthermore, the native quantile grid of Chronos-Bolt [12] ends at 0.9, so its violation rate cannot drop below 13.1% at any served quantile, while the heavy-tail mixture head of C ENTILE serves a calibrated τ =0.95 and reaches 4.7%. The served distribution is also the sharpest: The provisioning gain could in principle come from the operating point alone, so Table VII compares the served distribution itself against size-matched sequence baselines. We observe that C ENTILE attains the best MASE and CRPS at every horizon, 0.928 versus 0.958 MASE against the GRU at one hour, and the best or tied pinball loss at the 0.99 quantile. This can be attributed to the mixture head, which places explicit mass on the rare traffic spikes that a squared-error head averages away, and the advantage widens with the horizon. These results underscore that the same design serves a second decision in a second domain without modification. E. Ablation: Where the Gains Come From The choice of operating point is robust: Sweeping the served quantile from q=0.5 to q=0.8 leaves C ENTILE below both the deployed estimates and the history baseline at every setting on both primary capacities, and at q=0.5 and q=0.6 it outperforms every estimator. Together with the measured

8

Table VII. Forecast quality on CESNET against size-matched sequence models trained identically on the same streams, where GRU is a gated recurrent unit and TGR a gated-retention model (smaller values indicate better performance). MASE

Pinball@0.99

Table VIII. Deployment profile of the C ENTILE runtime estimator. All timing and size entries are measured.

CRPS

h=1 h=4 h=24 h=1 h=4 h=24 h=1 h=4 h=24

Model

GRU 0.958 1.079 1.086 0.031 0.037 0.035 0.561 0.631 0.623 TGR 1.106 1.278 1.115 0.037 0.046 0.043 0.637 0.717 0.655 C ENTILE 0.928 0.946 1.034 0.029 0.033 0.035 0.540 0.550 0.587 Empirical coverage

0.9 0.8 0.7 0.6 0.5 0.6

0.7

0.8

Measured value

Parameters Artifact size (bf16) Training, one A30 GPU Per-job CPU inference (4 cores) New-month deployment

1.32 M 2.5 MiB approx. 5 minutes 4.2 ms batched, 7.7 ms single zero-shot, no retraining

adapted billion-parameter models [11]. Cold-user estimates degrade gracefully, clipping to the declared limit and affecting only 0.8% of submissions with fewer than three prior observations. Nothing in the design is specific to these two decisions, as any quantile-consuming policy inherits the same calibration and inference path. In essence, a telemetry model’s value is revealed where it is deployed, and C ENTILE is designed to be evaluated there.

1.0

0.5

Property

0.9

Nominal quantile q

V. R ELATED W ORK Foundation models for networking and systems: Pretrained traffic models such as ET-BERT [7], netFound [8], NetGPT [34], and NetMamba [35] learn packet and flow representations, mainly for classification, while NetLLM [11] adapts a frozen language model to networking tasks and MobiGPT [9], UoMo [14], and LWM [10] target wireless networks. Each remains within a single domain. Time-series forecasting and foundation models: Deep forecasters are trained per dataset and per horizon, whereas Chronos [12], TimesFM [13], Moirai [36], Lag-Llama [37], and MOMENT [38] pretrain once on broad corpora, and another line reprograms frozen language models into forecasters [39], [40]. Closest to our setting, the Tiny Time Mixer was evaluated zero-shot on CESNET traffic [41], [42], CLOUDADV aligns zero-shot forecasts with cloud instance sizing [43], and decision-focused fine-tuning lowers feederdispatch cost [44]. Yet, each treats inputs as uniformlysampled series and evaluates one decision in one domain. Because estimate accuracy governs backfilling performance, predictors draw on system-generated recent-history predictions [18], machine learning [32], [45], uncertainty-aware runtime prediction [33], and learned backfilling policies [46], all retrained per system and exposing a point estimate. Learning over systems telemetry: Machine learning also drives operations tasks, from workload forecasting [1], [2] and autoscaling [47] to telemetry imputation [48], traffic engineering [49], and system log representation [50], but one task-specific model at a time. Network digital twins evaluate control policies on a mirrored network before deployment [51], drive data-driven optimization of reliable edge caching [52], and revise the twin’s ingested state as telemetry changes [53], with their creation and optimization surveyed in [54]. In situ video-streaming experiments reached the same conclusion, as proxy-metric gains failed to carry to the deployed decision [55]. C ENTILE instead pretrains one generative model on systems telemetry, replacing the per-task fleet with one design evaluated at the decisions it serves.

Fig. 5. Nominal quantile vs. empirical coverage of the runtime estimator against a protocol-matched GRU.

calibration of Fig. 5, this robustness lets the served quantile be chosen from the decision’s risk preference alone. Direct decoding outperforms rollout at an order of magnitude lower cost: With identical weights on the size-matched gated-retention model, rollout error compounds to 8.34 MASE by 64 steps while direct decoding holds 4.35, the gap widening monotonically with horizon in every seed. One direct pass serves all 25 input-length and horizon combinations in 5.2 seconds, versus 69.4 seconds for one rollout combination. The sequence is the signal: Both remaining ablations use a covariate-matched three-seed protocol, so their numbers fall slightly below the five-seed results of Table III. Removing the user’s event history degrades the estimator past the operator status quo (126.9 versus 94.8 at 20K nodes), while removing the current job’s covariates leaves it nearly unchanged (20.2 versus 20.0), so the per-user sequence carries nearly all of the decision-relevant signal. The attention layers are necessary: Under this protocol, the C ENTILE attention layers reach a mean bounded slowdown of 20.0 ± 1.4 and 15.5 ± 0.3 at the two primary capacities. A size-matched GRU trained identically reaches 27.4 ± 4.2 and 20.7 ± 2.7, leaving C ENTILE roughly a quarter lower with a three- to ninefold smaller seed standard deviation. Figure 5 explains the gap, as the nominal quantiles of C ENTILE track their empirical coverage with a mean gap of 0.07 versus the GRU’s 0.17. F. Deployment Considerations One artifact to operate: Table VIII profiles the estimator, which adds negligible cost to a scheduling cycle operating at second granularity, with no GPU on the inference path. A specialist approach maintains one model per task, per horizon, and per period, while C ENTILE collapses this product into one 2.5 MiB artifact per domain, three orders of magnitude below

9

VI. C ONCLUSION

[21] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” in NeurIPS, 2017. [22] J. Zhai et al., “Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations,” in ICML, 2024. [23] Z. Hou, W. Gao, Y. Shen, F. Wang, and X. Liu, “Protransformer: Robustify transformers via plug-and-play paradigm,” in NeurIPS, 2024. [24] H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. Jégou, “Going deeper with image transformers,” in ICCV, 2021. [25] C. M. Bishop, “Mixture density networks,” Aston University, Tech. Rep., 1994. [26] R. Wen, K. Torkkola, B. Narayanaswamy, and D. Madeka, “A multihorizon quantile recurrent forecaster,” arXiv:1711.11053, 2017. [27] A. Venkatraman, M. Hebert, and J. A. Bagnell, “Improving multi-step prediction of learned time series models,” in AAAI, 2015. [28] H. Hewamalage, C. Bergmeir, and K. Bandara, “Recurrent neural networks for time series forecasting: Current status and future directions,” Int. J. Forecasting, 2021. [29] G.-Y. Ban and C. Rudin, “The big data newsvendor: Practical insights from machine learning,” Oper. Res., 2019. [30] J. Koumar, K. Hynek, T. Čejka, and P. Šiška, “Cesnet-timeseries24: Time series dataset for network traffic anomaly detection and forecasting,” Scientific Data, 2025. [31] S. Makridakis, E. Spiliotis, and V. Assimakopoulos, “The M4 competition: 100,000 time series and 61 forecasting methods,” Int. J. Forecasting, 2020. [32] E. Gaussier, D. Glesser, V. Reis, and D. Trystram, “Improving backfilling by using machine learning to predict running times,” in SC, 2015. [33] J. Choi and S. Oh, “Uarp: Uncertainty-aware runtime prediction for preventing scheduler termination under wallclock constraints in hpc,” J. Supercomput., 2026. [34] X. Meng, C. Lin, Y. Wang, and Y. Zhang, “Netgpt: Generative pretrained transformer for network traffic,” arXiv:2304.09513, 2023. [35] T. Wang, X. Xie, W. Wang, C. Wang, Y. Zhao, and Y. Cui, “NetMamba: Efficient network traffic classification via pre-training unidirectional mamba,” in IEEE ICNP, 2024. [36] G. Woo, C. Liu, A. Kumar, C. Xiong, S. Savarese, and D. Sahoo, “Unified training of universal time series forecasting transformers,” in ICML, 2024. [37] K. Rasul et al., “Lag-llama: Towards foundation models for probabilistic time series forecasting,” arXiv:2310.08278, 2023. [38] M. Goswami, K. Szafer, A. Choudhry, Y. Cai, S. Li, and A. Dubrawski, “MOMENT: A family of open time-series foundation models,” in ICML, 2024. [39] M. Jin, S. Wang, L. Ma, Z. Chu, J. Y. Zhang, X. Shi, P.-Y. Chen, Y. Liang, Y.-F. Li, S. Pan, and Q. Wen, “Time-LLM: Time series forecasting by reprogramming large language models,” in ICLR, 2024. [40] T. Zhou, P. Niu, X. Wang, L. Sun, and R. Jin, “One fits all: Power general time series analysis by pretrained LM,” in NeurIPS, 2023. [41] V. Ekambaram, A. Jati, P. Dayama, S. Mukherjee, N. H. Nguyen, W. M. Gifford, C. Reddy, and J. Kalagnanam, “Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,” arXiv:2401.03955, 2024. [42] F. Liu, B. Farkiani, and P. Crowley, “Time-series foundation models for isp traffic forecasting,” arXiv:2511.17529, 2025. [43] J. Bell, G. Carfi, G. Gramaglia, A. Simioni, D. Fontani, and V. Lomonaco, “Cloudadv: Decision-aligned instance sizing with zeroshot foundation models under drift,” arXiv:2606.31470, 2026. [44] M. Beichter, N. Friederich, J. Pinter, D. Werling, K. Phipps, S. Beichter, O. Neumann, R. Mikut, V. Hagenmeyer, and B. Heidrich, “Decisionfocused fine-tuning of time series foundation models for dispatchable feeder optimization,” Energy and AI, 2025. [45] D. Klusáček and M. Soysal, “Walltime prediction and its impact on job scheduling performance and predictability,” in JSSPP, 2020. [46] E. Kolker-Hicks, D. Zhang, and D. Dai, “A reinforcement learning based backfilling strategy for hpc batch jobs,” in SC Workshops (SC-W), 2023. [47] K. Rzadca, P. Findeisen, J. Swiderski, P. Zych, P. Broniek, J. Kusmierek, P. Nowak, B. Strack, P. Witusowski, S. Hand, and J. Wilkes, “Autopilot: Workload autoscaling at Google,” in ACM EuroSys, 2020. [48] F. Gong, D. Raghunathan, A. Gupta, and M. Apostolaki, “Zoom2Net: Constrained network telemetry imputation,” in ACM SIGCOMM, 2024. [49] Y. Perry, F. V. Frujeri, C. Hoch, S. Kandula, I. Menache, M. Schapira, and A. Tamar, “DOTE: Rethinking (predictive) WAN traffic engineering,” in USENIX NSDI, 2023.

In this paper, we introduced C ENTILE, a generative foundation model for network and systems telemetry evaluated at the level of the decisions it drives. C ENTILE treats telemetry as irregularly timed entity streams and serves every horizon as calibrated quantiles in a single pass from one pretrained model per domain. Extensive experiments on Fugaku job logs and national ISP traffic confirm that this design lowers the mean bounded slowdown of backfilling by up to approximately 77% over deployed user estimates and roughly halves the deployed rule’s violation rate. The 2.5 MiB runtime estimator transfers zero-shot across months and serves in milliseconds on commodity CPUs. VII. U SE OF AI D ISCLOSURE AI tools [56] supported code development and manuscript editing, and the authors reviewed all generated content for accuracy. R EFERENCES [1] E. Cortez, A. Bonde, A. Muzio, M. Russinovich, M. Fontoura, and R. Bianchini, “Resource central: Understanding and predicting workloads for improved resource management in large cloud platforms,” in ACM SOSP, 2017. [2] A. Borghesi et al., “M100 exadata: a data collection campaign on the cineca’s marconi100 tier-0 supercomputer,” Scientific Data, 2023. [3] C. Reiss, J. Wilkes, and J. L. Hellerstein, “Google cluster-usage traces: Format + schema,” Google Inc., Tech. Rep., 2011. [4] M. Tirmazi et al., “Borg: The next generation,” in EuroSys, 2020. [5] A. Paleyes, R.-G. Urma, and N. D. Lawrence, “Challenges in deploying machine learning: A survey of case studies,” ACM Comput. Surv., 2022. [6] J. Mandi, J. Kotary, S. Berden, M. Mulamba, V. Bucarey, T. Guns, and F. Fioretto, “Decision-focused learning: Foundations, state of the art, benchmark and future opportunities,” J. Artif. Intell. Res., 2024. [7] X. Lin, G. Xiong, G. Gou, Z. Li, J. Shi, and J. Yu, “Et-bert: A contextualized datagram representation with pre-training transformers for encrypted traffic classification,” in WWW, 2022. [8] S. Guthula, N. Battula, R. Beltiukov, W. Guo, and A. Gupta, “netfound: Foundation model for network security and performance,” arXiv:2310.17025, 2023. [9] X. Qi, H. Chai, and Y. Li, “Mobigpt: A foundation model for mobile wireless networks,” arXiv:2509.18166, 2025. [10] S. Alikhani, G. Charan, and A. Alkhateeb, “Large wireless model (lwm): A foundation model for wireless channels,” arXiv:2411.08872, 2024. [11] D. Wu, X. Wang, Y. Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “Netllm: Adapting large language models for networking,” in ACM SIGCOMM, 2024. [12] A. F. Ansari et al., “Chronos: Learning the language of time series,” TMLR, 2024. [13] A. Das, W. Kong, R. Sen, and Y. Zhou, “A decoder-only foundation model for time-series forecasting,” in ICML, 2024. [14] H. Chai, S. Zhang, X. Qi, B. Qiu, and Y. Li, “Uomo: A universal model of mobile traffic forecasting for wireless network optimization,” in ACM KDD, 2025. [15] F. Antici, A. Bartolini, J. Domke, Z. Kiziltan, and K. Yamamoto, “Fdata: A fugaku workload dataset for job-centric predictive modelling in hpc systems,” Scientific Data, 2025. [16] J. Guo, Z. Chang, S. Wang, H. Ding, Y. Feng, L. Mao, and Y. Bao, “Who limits the resource efficiency of my datacenter: An analysis of alibaba datacenter traces,” in IEEE/ACM IWQoS, 2019. [17] D. A. Lifka, “The ANL/IBM SP scheduling system,” in JSSPP, 1995. [18] D. Tsafrir, Y. Etsion, and D. G. Feitelson, “Backfilling using systemgenerated predictions rather than user runtime estimates,” IEEE TPDS, 2007. [19] D. Carastan-Santos and R. Y. de Camargo, “Obtaining dynamic scheduling policies with simulation and machine learning,” in SC, 2017. [20] T. Gneiting and A. E. Raftery, “Strictly proper scoring rules, prediction, and estimation,” J. Amer. Statist. Assoc., 2007.

10

[50] Z. Hou, M. Ghashami, M. Kuznetsov, and M. Torkamani, “Hlogformer: A hierarchical transformer for representing log data,” arXiv:2408.16803, 2024. [51] M. Hossen, Z. Zhang, D. Ron, Y. Liu, and V. K. Shah, “OpenTwin: Digital twin driven closed-loop KPM inference and control for open RAN,” arXiv, 2026. [52] Z. Zhang, Y. Liu, Z. Peng, M. Chen, D. Xu, and S. Cui, “Digital twinassisted data-driven optimization for reliable edge caching in wireless networks,” IEEE JSAC, 2024. [53] Z. Zhang, D. Chen, A. Gao, M. Wang, M. Chen, M. Fang, X. Yang, and Y. Liu, “Network digital untwinning: Towards backward optimization of digital twins,” in IEEE ICDCS, 2026. [54] Z. Zhang, Z. Peng, H. Yu, M. Chen, and Y. Liu, “Digital network twins for next-generation wireless: Creation, optimization, and challenges,” IEEE Network, 2025. [55] F. Y. Yan, H. Ayers, C. Zhu, S. Fouladi, J. Hong, K. Zhang, P. Levis, and K. Winstein, “Learning in situ: A randomized experiment in video streaming,” in USENIX NSDI, 2020. [56] Anthropic, “Claude,” https://www.anthropic.com/claude, 2026.

11

Related documents

• My Cosmos: development, clinical protocol, and preliminary usability of a gamified transdiagnostic digital CBT platform.#138421
• Physicians’ acceptance of large language model–based clinical decision support tools in gynecologic oncology: a technology acceptance model study.#409356
• Prospective Registration of mOrbidity and Mortality, individUal radioSensitivity and Radiation Technique (PROUST)#1137570
• Comparison and Opinion on the Safety of Teleconsultation Versus Traditional Anaesthetic Consultation in a Day Hospital#800533
• Performance evaluation of automated medicine delivery systems#137525