PaperlayerView exact revision on arXiv

Source-linked AI summary

Capturing Electrical Windows Through Reconfigurable Optical Networks: Transfer-aware Scheduling of Multi-Region AI Data Centers

Xi Tong, Zijian Li, Xiaoyi Fan, Cong Zhang, Ershun Du, Qin Wang

arXiv:2610.13061v1cs.DC

TL;DR

AIDCs can migrate training to follow regional electricity prices and renewable-energy availability, but checkpoint delays can leave favorable windows unreachable. This paper coordinates workload placement with checkpoint delivery and optical connectivity; in a five-AIDC benchmark, Dynamic OCS lowers grid-purchase cost and raises renewable utilization relative to static optical operation.

  • Problem

    The coupling between network-constrained migration and electrical-window capture remains insufficiently explored, and transfer delays may prevent training from exploiting short-lived favorable windows.

    Source: C. Research Gaps and Motivations · p. 2

  • Method

    The framework links workload placement and execution timing to checkpoint delivery and reconfigurable optical connectivity.

    Source: §VI · p. 9

  • Results

    12.50% lower RT grid-purchase cost and 2.724 percentage points higher RE utilization with Dynamic OCS versus Static optical in the controlled five-AIDC benchmark.

    Source: §VI · p. 10

  • Takeaways & Limitations

    Sensitivity and ablation analyses indicate that OCS gains depend on electrical opportunities aligning with network bottlenecks that constrain timely checkpoint delivery.

    Source: E. Sensitivity and Ablation Analysis · p. 9, §VI · p. 10

Abstract

from arXiv · show

Artificial-intelligence data centers (AIDCs) can provide spatiotemporal load flexibility by migrating training workloads in response to regional electricity prices and renewable-energy (RE) availability. However, transfer-blind scheduling may overestimate migration benefits by overlooking transfer time or prevent training from resuming within renewable-rich or low-price windows when considering transfer time. This paper proposes a transfer-aware compute-electricity-network scheduling framework that coordinates training placement, checkpoint transmission, and optical circuit switching (OCS) to improve local RE utilization and reduce grid-purchase cost. A compact compute-electricity master problem generates candidate schedules using estimated transfer delays. The framework then schedules optical circuits, verifies checkpoint delivery under shared-link capacity constraints, and revises infeasible schedules using updated delay estimates. It integrates day-ahead planning, event-triggered intraday updates, and real-time dispatch. In a five-AIDC benchmark with equal completed work and network resources, the proposed method reduces the real-time grid purchase cost by 12.50% and increases the local RE utilization by 2.724 percentage points relative to static optical operation. Sensitivity and ablation analyses indicate that OCS is particularly beneficial when renewable-rich or low-price windows shift across sites and network bottlenecks hinder timely checkpoint delivery.

I. INTRODUCTION

AIDCs can shift training across time and regions to respond to electricity prices and renewable-energy availability, but checkpoint-transfer delays can make favorable electrical windows unreachable. The paper links workload scheduling with network capacity and electrical opportunities to address this gap.

I. INTRODUCTION

  • AIDCs offer both temporal and spatial flexibility by shifting computing work across time and regions in response to electricity prices and renewable-energy availability.Unlike conventional flexible resources that mainly shift consumption in time, AIDCs can also relocate work across regions.Source: A. Background and Motivation · p. 1
  • Transfer-blind scheduling can overestimate migration benefits because delayed checkpoint delivery may make an electrically attractive destination unreachable before its favorable window closes.Inter-site bandwidth bottlenecks can further delay delivery when migration demand shifts between AIDC pairs.Source: A. Background and Motivation · pp. 1–2

B. Literature Review

  • Prior power-system studies value workload flexibility while networking studies optimize communication, but neither makes time-varying inter-site capacity determine electrical-window reachability.Earlier formulations either abstracted communication or treated transmission speed, distance, latency, or bandwidth as prescribed parameters.Source: B. Literature Review · p. 2

C. Research Gaps and Motivations

  • The paper proposes a transfer-aware compute–electricity–network framework that coordinates workload migration, network transmission, and time-varying electrical opportunities.It incorporates OCS into a coupled model with checkpoint precedence, inference-priority headroom, and network-resource constraints.Source: C. Research Gaps and Motivations · p. 2
  • The method combines day-ahead, event-triggered intraday, and one-minute real-time scheduling, using transfer-lag feedback and network certification to preserve migration executability.The paper terms the overlooked network factor “transfer-blind” and defines “electrical-window reachability” to assess the gap between current and expected transfer modes.Source: C. Research Gaps and Motivations · pp. 2–3

II. SYSTEM DESCRIPTION AND SCHEDULING FRAMEWORK

The system models geographically separated AIDCs with local electricity and renewable-energy conditions, inference demand, and checkpointed training. Its framework treats checkpoint delivery and reconfigurable optical connectivity as determinants of whether training can resume within an electrical opportunity window.

A. Multi-Region AIDC and Network Description

  • The modeled portfolio comprises geographically separated AIDCs, each with a regional grid interface, time-varying electricity price, and local renewable generation.Facility demand includes base power, priority inference, and checkpointed training; the grid supplies residual demand and curtailment is unused available local renewable energy.Source: A. Multi-Region AIDC and Network… · p. 3

B. Checkpoint Migration and Matched Network Resources

  • Training jobs consist of ordered compute segments, and a successor at another site cannot start until its predecessor finishes and the checkpoint is packed, transferred, and restored.The first segment must run at the job's origin, while later segments may run at eligible AIDCs.Source: B. Checkpoint Migration and Matched… · p. 3
  • Dynamic OCS redirects or re-pairs the same installed circuits as transfer demand changes, redistributing rather than adding communication capacity.Static optical operation keeps circuit pairings and directions fixed.Source: B. Checkpoint Migration and Matched… · p. 3
  • Electrical-window reachability requires packing, transfer, and restoration to finish early enough for successor execution before the attractive site–time opportunity closes.Under matched resources, Dynamic OCS can redirect capacity toward a binding migration so the successor is ready while the same renewable-rich opportunity remains available.Source: B. Checkpoint Migration and Matched… · pp. 3–4

C. Overall Scheduling Framework

The framework links electrical opportunities and checkpoint delivery through successor-readiness timing, then coordinates day-ahead, intraday, and real-time scheduling. Compute scheduling places ordered training segments and delays successors when migration requires transfer time.

C. Overall Scheduling Framework

  • Day-ahead scheduling sets the baseline, intraday scheduling revises uncommitted decisions when updated information matters, and real-time scheduling executes the latest feasible plan at one-minute resolution.Successor-readiness timing couples checkpoint delivery to destination load.Source: C. Overall Scheduling Framework · p. 4

A. Compute Scheduling

  • Each job is divided into ordered, checkpoint-separated segments, with the first segment fixed at its prescribed origin and later segments assigned admissible site–time starts.The model converts starts into slot-level GPU occupancy and reserves safe capacity after serving nondeferrable inference.Source: A. Compute Scheduling · p. 4
  • For each transition, the model identifies predecessor and successor sites and delays a successor for cross-site checkpoint transfer using a screened inter-site lag.The lag is certified in the scheduling procedure.Source: A. Compute Scheduling · p. 4

B. Electricity Interface

The electricity interface maps GPU occupancy to facility demand and accounts for local renewable energy, grid purchases, and facility limits. Forecast handling also considers inference demand and electrical exposure under day-ahead and intraday forecasts.

B. Electricity Interface

  • Occupied GPUs are mapped to IT power and then multiplied by site PUE to represent facility demand.The formulation identifies base IT power, incremental power per active GPU, and site PUE.Source: B. Electricity Interface · p. 4
  • Facility demand is supplied by local renewable energy and the grid, with equations defining grid purchase, renewable use, curtailment, and a facility limit.The balance uses local renewable availability.Source: B. Electricity Interface · pp. 4–5
  • Peak and ramp exposure must remain feasible under both day-ahead and intraday forecasts, using the larger inference forecast.The renewable forecast treatment introduces an epigraph represented by two linear lower bounds and a small positive tie-break, avoiding an additional binary variable.Source: B. Electricity Interface · p. 5

C. Shared Network and OCS

Checkpoint flows are scheduled across paths and time to meet compute-created transfer windows while sharing link capacity. The network model combines packet and optical capacity and represents optical-circuit reconfiguration within deployment-specific budgets.

C. Shared Network and OCS

  • Checkpoint volumes are scheduled over paths and time so each transfer arrives within its compute-created window while concurrent flows share link capacity.Candidate paths have at most three hops in this study; the delivery constraint covers each checkpoint size and the capacity constraint couples concurrent migrations.Source: C. Shared Network and OCS · p. 5
  • Effective transfer capacity combines residual packet capacity with active optical capacity, accounting for background traffic and reconfiguration loss.The model uses a conversion factor for rate-times-time volume and tracks optical circuits and new activations.Source: C. Shared Network and OCS · p. 5
  • Dynamic OCS optimizes circuit assignments under the same circuit and site-port budgets as static optical operation, while enforcing activation detection and minimum hold time.Static optical fixes circuit assignments; dynamic OCS may re-pair or redirect installed circuits as transfer demand changes, redistributing rather than adding communication capacity.Source: C. Shared Network and OCS · p. 5

D. Objective Function and Constraints

The objective combines energy, curtailment, network, peak, imbalance, and grid-behavior costs. It measures aggregate grid purchases and ramping across forecast vintages, while keeping AC-network constraints outside the site-level grid-interface boundary.

D. Objective Function and Constraints

  • Grid behavior is measured through worst aggregate purchase, total ramping, and maximum ramping under both forecast vintages.The objective also penalizes energy use and undesirable grid behavior.Source: D. Objective Function and Constraints · p. 5
  • The objective combines energy, curtailment, migration, switching, peak, ramp, and intraday-deviation costs.The deviation term penalizes intraday departures from the day-ahead aggregate schedule and is zero in day-ahead planning.Source: D. Objective Function and Constraints · p. 5
  • AC-network constraints remain outside the site-level grid-interface boundary, and all cases share the same coefficients.Source: D. Objective Function and Constraints · p. 5

IV. THREE-STAGE SCHEDULING AND NETWORK CERTIFICATION

The framework combines DA–ID rolling coordination with a compact compute-electricity master and exact network certification. It revises migration delays when checkpoint transfers fail certification.

IV. THREE-STAGE SCHEDULING AND NETWORK CERTIFICATION

  • Candidate generation and exact multicommodity-flow certification keep the master compact and produce network-feasible DA/ID plans over prescribed paths, without claiming global optimality.The repair loop warm-starts after each update and stops at its tolerance or repair limit.Source: §IV · p. 5, B. Compact Master and Exact… · p. 6

A. DA–ID Rolling Coordination

  • DA–ID coordination inherits committed starts and circuits, then uses changes in inference, renewable power, price, and background traffic to trigger replanning.Coarse windows retain the largest signal change, prioritizing those above the intraday threshold; the implementation uses a controlled forecast-error trajectory, while deployment would use streaming forecasts.Source: A. DA–ID Rolling Coordination · pp. 5–6

B. Compact Master and Exact Certification

  • The master represents each migration’s network time with a lag initialized conservatively from residual maximum-flow capacity.The estimate uses lower-quantile residual packet capacity and a safety factor; observed throughput enlarges only the corresponding binding lag after a failed certificate.Source: B. Compact Master and Exact… · p. 6
  • The compact master proposes compute plans, then a port-feasible OCS schedule and exact flow certification verify checkpoint delivery under shared-link constraints.Plans are accepted only within the network tolerance; failed migrations trigger updates to their binding lag estimates.Source: §IV · p. 5, B. Compact Master and Exact… · p. 6

C. Dynamic Candidate Portfolio and RT Dispatch

Dynamic OCS compares certified reconfigurable-network and static-reference candidates, retaining the lower-cost feasible plan. Real-time dispatch prioritizes urgent checkpoint transfers while respecting inference priority and successor-start conditions.

C. Dynamic Candidate Portfolio and RT Dispatch

  • The dynamic policy certifies candidates from reconfigurable-network and static-reference lags, then retains the lower-cost feasible candidate.Because Dynamic OCS can retain the static circuit, the certified Static plan serves as the fallback incumbent.Source: C. Dynamic Candidate Portfolio and RT… · p. 6
  • At one-minute resolution, RT dispatch gives checkpoint priority according to successor-start and job-deadline urgency.Every 15 minutes, Dynamic OCS scores port-feasible matchings by remaining volume, urgency, path length, and the certified plan; a new circuit loses its first minute.Source: C. Dynamic Candidate Portfolio and RT… · p. 6
  • Inference has priority, active segments are non-preemptive, and a successor requires completed compute, completed transfer, and available GPUs.The main experiments disable early starts, so real-time operation may delay but never advance the accepted plan.Source: C. Dynamic Candidate Portfolio and RT… · p. 6

V. CASE STUDY

The case study evaluates four scenarios in a five-AIDC benchmark with constructed workload and electricity inputs. At equal completed work, Dynamic OCS lowers grid-purchase cost and increases local renewable utilization relative to Static optical operation.

A. Scenario and Data Settings

  • The five geographically distributed AIDCs compare four scenarios; C and D both allow multi-hop routing and shared-link competition but differ in circuit direction.C fixes the circuit direction, while D allows it to be reconfigured.Source: A. Scenario and Data Settings · p. 6, Table I · p. 7
  • Training inputs use 16 Kalos pretraining records with 256–1024 GPUs and 2.09–13.75-hour durations; each job has at most four compute segments and two migrations.The benchmark assumes 2.10–12.15-TB job states and applies a network-pressure multiplier of 5.775–33.413 TB per migrated state.Source: A. Scenario and Data Settings · p. 6
  • Inference profiles derive from BurstGPT requests, while CAISO data and controlled site redistribution construct the electricity inputs.The controlled transformation shifts renewable availability and inference reservations across sites and adds locational prices while preserving each slot’s total renewable power, total inference GPUs, and mean price.Source: A. Scenario and Data Settings · p. 7, Figure 7 · p. 7

B. Comparison of Electricity Cost and Renewable Utilization

  • 12.50% lower cost for D than C and a 2.724-percentage-point higher RE utilization for D than C at equal completed work.All four cases consume 69.017 MWh of facility energy and finish without deadline violations, GPU overload, or incomplete segments; D uses 19.554 MWh of grid energy versus C’s 21.345 MWh.Source: Figure 8 · p. 7, B. Comparison of Electricity Cost and… · p. 7

C. Migration Scheduling and Renewable-Window Capture

Dynamic OCS shifts training across rotating site-level renewable windows by redirecting circuits toward changing transfer bottlenecks. It also reduces major grid-facing corrections, though total absolute ramping rises slightly.

C. Migration Scheduling and Renewable-Window Capture

  • D reallocates training across rotating site-level windows, yielding a 1.791 MWh final gain equal to the reduction in grid energy.D executes 20 transfers versus C’s 13, while mean duration falls from 69.31 to 17.10 min and successor-readiness lateness falls from 32 min to zero; the transfer populations differ, so duration curves are system-level rather than paired-transfer evidence.Source: C. Migration Scheduling and… · p. 8

D. Grid-Interaction Characteristics

  • D reduces sitewise DA–RT deviation by 29.19% and the maximum 15-min ramp by 28.55%, while the 15-min peak and P95 ramp fall by 4.15% and 5.56%.The stronger reductions in maximum ramp and schedule deviation indicate that Dynamic OCS primarily limits the most consequential grid-facing corrections.Source: D. Grid-Interaction Characteristics · p. 8, Figure 11 · p. 9, Table IV · p. 9
  • Total absolute ramp rises from 17.644 to 17.834 MW (1.08%), so Dynamic OCS reshapes grid imports rather than uniformly smoothing them.The exceedance curves stay close at lower magnitudes and separate in the upper tail, with lower peak exposure and closer DA schedule tracking.Source: D. Grid-Interaction Characteristics · pp. 8–9

E. Sensitivity and Ablation Analysis

Dynamic OCS produces benchmark gains, with sensitivity and ablation analyses showing that benefits depend on spatial electrical opportunities coinciding with network bottlenecks. The conclusions are bounded by constructed inputs and a site-level grid-interface model.

E. Sensitivity and Ablation Analysis

  • At every spatial intensity, raising network pressure from 2.0 to 3.0 increases cost reduction by 8.41–10.77 points and RE-utilization gain by 1.77–2.88 points.Across the screen, the largest cost benefit is 15.76% at intensity 0.5, while the largest RE gain is 3.85 points at intensity 1.0; interior cells are not monotonic.Source: E. Sensitivity and Ablation Analysis · p. 9
  • With no added network pressure, price or renewable redistribution leaves cost differences between −0.52% and 0.02% and RE-utilization differences between 0 and 0.08 points.Network pressure alone yields 4.20% and 0.71 points, while all three drivers yield 10.05% and 2.28 points; clear gains emerge when electrical opportunity coincides with a constraining transfer network.Source: E. Sensitivity and Ablation Analysis · p. 9
  • Eight of 12 screening points deliver double-digit cost reductions, and five show RE-utilization gains above two points; no pressure-2.0 point reaches either threshold.A fixed circuit is nearly sufficient under weak transfer pressure, while Dynamic OCS creates clear value when required migration directions change across short-lived price and renewable windows.Source: E. Sensitivity and Ablation Analysis · p. 9

VI. CONCLUSION

  • The framework links workload placement and execution timing to checkpoint delivery and reconfigurable optical connectivity.Under matched optical resources, Dynamic OCS redirects capacity toward changing transfer bottlenecks, improving access to local renewable windows.Source: §VI · p. 9
  • 12.50% lower RT grid-purchase cost and 2.724 percentage points higher RE utilization versus Static optical in the controlled five-AIDC benchmark.The reported gains depend on alignment between spatial electrical opportunities and network bottlenecks; peak import, maximum ramp, and DA–RT deviations decrease, while cumulative absolute ramping increases slightly.Source: §VI · p. 10
  • The findings are bounded by constructed checkpoint sizes, transformed site profiles, and a site-level grid-interface model without AC-network or reserve-market constraints.Source: §VI · p. 10
Loading 2610.13061v1…