PaperlayerView exact revision on arXiv

Source-linked AI summary

Grid Demand Flexibility Assessment of AI Data Centers via Batch Workload Temporal Shifting

Suntao Su, Liang Du, Shengyi Wang

arXiv:2609.38020v1eess.SYcs.SY

TL;DR

AI data centers have increasingly significant and variable power demand, while existing flexibility assessments do not fully address workload-to-grid timing mismatch or combined CPU, GPU, and memory utilization. This paper processes fine-resolution traces into grid-compatible intervals and models batch shifting with workload and server constraints. Its results quantify peak-shaving flexibility and show a tradeoff between sustained reduction magnitude and duration.

  • Problem

    Existing assessments do not sufficiently address the mismatch between fine-grained workload traces and grid-operation intervals or fully characterize the combined effects of CPU, GPU, and memory use.

    Source: §I · p. 1

  • Method

    The framework maps fine-resolution CPU, GPU, and memory traces to grid intervals, schedules batch workloads within timing and server constraints, and estimates power using a utilization-dependent server model.

    Source: §I · pp. 1–2, A. Potential Temporal Flexibility in… · p. 2, B. Deferrable Jobs · p. 2

  • Results

    In Case 2, flexibility increases from 10.76 kW at γ = 0 to 32.47 kW at γ = 1.0, while sustained reductions show a magnitude–duration tradeoff.

    Source: B. Peak Power Demand Shaving · p. 4, C. Sustained Power Demand Reduction · p. 5

  • Takeaways & Limitations

    The framework provides a means to quantify both the magnitude and duration of AI data-center demand flexibility for power system operation.

    Source: §VI · p. 5

Abstract

from arXiv · show

The rapid growth of artificial intelligence (AI) data centers has introduced new challenges to power system operation. As their power demand becomes larger and more variable, quantitatively characterizing their demand flexibility is increasingly important for effective power system coordination. However, heterogeneous workload characteristics and resource requirements make this flexibility difficult to characterize directly. This paper proposes a framework for assessing the grid-compatible demand flexibility of AI data centers via batch workload temporal shifting. An averaging-based resource usage processing method is developed to map fine-resolution CPU, GPU and memory usage into unified time intervals compatible with power system operation. A workload temporal scheduling model is then formulated to shift batch workloads while preserving execution continuity, delay constraints, and server resource capacities, and is coupled with a utilization-dependent server power model to translate workload scheduling decisions into server power demand. Two complementary flexibility metrics are evaluated: short-term peak demand shaving and the maximum duration of sustained power reduction. Numerical results based on real GPU cluster traces demonstrate that workload temporal shifting can provide quantifiable and grid-compatible demand flexibility for AI data centers with limited disruption to computing workloads.

I. INTRODUCTION

AI data centers create time-varying power demand, yet existing flexibility assessments do not fully address workload-to-grid time-resolution mismatch or the combined role of CPU, GPU, and memory use. The paper proposes a framework to assess grid-compatible flexibility through batch workload shifting.

I. INTRODUCTION

  • AI data centers are both a growing, variable power demand and a potential flexible demand-side resource through workload shifting.Prior work cited in the introduction reports software orchestration achieving a 25% power reduction for three hours while maintaining workload quality of service [4],.Source: §I · p. 1
  • Existing assessments leave unresolved the mismatch between fine-grained workload traces and grid-operation intervals, and do not fully characterize combined CPU, GPU, and memory effects.These limitations constrain direct use in grid scheduling and control.Source: §I · p. 1
  • The framework maps fine-resolution resource traces into grid-compatible intervals and integrates batch scheduling, server resource constraints, and utilization-dependent power estimates.It translates workload scheduling decisions directly into power consumption.Source: §I · pp. 1–2

II. MODULAR DATA OF WORKLOADS

The section distinguishes shiftable offline batch jobs from delay-sensitive online jobs, then defines a grid-aligned interval scheme for identifying deferrable workloads. Batch flexibility comes from waiting time between submission and execution, bounded by job-specific timing limits.

A. Potential Temporal Flexibility in Workloads

  • Offline jobs such as AI training can tolerate waiting, whereas online jobs are delay-sensitive; only batch-workload flexibility is quantified.Batch jobs can have a waiting interval between submission and execution that provides an allowable shifting window.Source: A. Potential Temporal Flexibility in… · p. 2

B. Deferrable Jobs

  • Grid scheduling commonly uses 5-, 15-, or 60-minute intervals, unlike the workload dataset's second-level resolution, so workload data are transformed to grid-compatible intervals.The delay rate γ can extend a batch job's allowable shift by up to γD_i intervals beyond its original window.Source: B. Deferrable Jobs · p. 2
  • A job is deferrable when execution fits within the daily scheduling horizon and its allowable shifting interval is at least one grid interval.The section uses uniform intervals of duration δ, with T = (24 × 60)/δ intervals per day.Source: B. Deferrable Jobs · p. 2

C. Averaging-based Resource Usage Processing Method

The averaging-based processing method maps instance-level CPU, GPU, and memory use onto the scheduling intervals. It accounts for partial execution overlap within an interval and aggregates overlapping instances of the same job.

C. Averaging-based Resource Usage Processing Method

  • The processing covers CPU, GPU, and memory resource types.The resource set is denoted RE = {CPU, GPU, memory}.Source: C. Averaging-based Resource Usage… · p. 2
  • For each instance, the method computes overlap with each scheduling interval and uses that overlap to derive interval-level average resource usage.The resource usage attributed to an instance is assumed constant during execution and equal to its scheduler allocation.Source: C. Averaging-based Resource Usage… · p. 2
  • When multiple instances of one job overlap the same interval, their averaged resource usage is aggregated to obtain the job's interval-level usage.Source: C. Averaging-based Resource Usage… · p. 2

III. FLEXIBILITY MODEL OF BATCH WORKLOADS

The flexibility model combines batch-workload scheduling with a data-center IT power model. Its scheduling formulation respects workload execution continuity and feasible start times while accounting for server resource capacity.

III. FLEXIBILITY MODEL OF BATCH WORKLOADS

  • The model combines batch-workload shifting with data-center IT power consumption estimation.Source: §III · p. 2

A. Batch Workloads Shifting Model

  • Each batch job selects exactly one feasible starting time, represented by a binary schedule variable.The formulation reflects non-preemptive execution: a workload cannot be interrupted after it starts.Source: A. Batch Workloads Shifting Model · p. 3

B. Power Consumption Model of Servers

The server power model estimates workload-level electrical demand from resource utilization, using CPU, GPU, and memory rather than CPU alone. It assumes a predefined job-scheduling algorithm and accounts for online and offline workloads.

B. Power Consumption Model of Servers

  • The model assumes the scheduler uses a predefined algorithm to assign online and offline workload resource demands to servers with sufficient available resources.Assigned job resource demands are weighted by server and relative execution interval.Source: B. Power Consumption Model of Servers · p. 3
  • Scheduling decisions must keep each server’s CPU, GPU, and memory usage within its corresponding resource capacities.Source: B. Power Consumption Model of Servers · p. 3
  • Server power is estimated as a workload-level quantity using a linear model that relates power to CPU, GPU, and memory utilization, rather than as a hardware-accurate estimate.The model includes server idle power and rated resource power.Source: B. Power Consumption Model of Servers · p. 3

IV. GRID DEMAND FLEXIBILITY ASSESSMENT FRAMEWORK

The framework evaluates short-term peak demand shaving and sustained power reduction. For peak shaving, it maximizes guaranteed flexibility while minimizing the number of shifted batch jobs as a secondary objective.

IV. GRID DEMAND FLEXIBILITY ASSESSMENT FRAMEWORK

  • The framework assesses two operating capabilities: short-term peak demand shaving and sustained power reduction.Source: §IV · p. 3

A. Short-Term Peak Demand Shaving

  • The peak-shaving objective schedules the minimum number of batch workloads to quantify maximum grid demand flexibility during a critical high-demand period.Source: A. Short-Term Peak Demand Shaving · p. 3
  • Peak shaving maximizes guaranteed grid-compatible flexibility during a requested event while minimizing the number of shifted jobs as a secondary objective.The trade-off coefficient α is set to 0.01 so maximizing flexibility remains dominant.Source: A. Short-Term Peak Demand Shaving · p. 3

B. Sustained Power Reduction

Sustained power reduction assessment fixes the requested reduction magnitude and optimizes how long it can be maintained. The selected intervals must form one continuous flexibility period.

B. Sustained Power Reduction

  • The sustained-duration assessment fixes the requested reduction magnitude at F_req and optimizes the duration of the reduction.Source: B. Sustained Power Reduction · p. 3
  • Binary variables identify intervals meeting the requested reduction and the start of a sustained period, enforcing one continuous flexibility block over the scheduling horizon.Constraints (18)–(20) use a linear big-M constraint and restrict the horizon to one flexibility period.Source: B. Sustained Power Reduction · pp. 3–4

V. NUMERICAL RESULTS

The numerical experiments use one day of Alibaba Cluster Trace GPU v2020, comprising 7,828 jobs.

V. NUMERICAL RESULTS

  • The simulations use one day of Alibaba Cluster Trace GPU v2020, containing 7,828 jobs.They are implemented in Python with Jupyter 6.3.0 and Gurobi 13.0.3.Source: §V · p. 4

A. Temporal and Power Characteristics of AI Workloads

Batch workloads have greater temporal flexibility than online workloads, and temporal shifting can shave short-term peaks without directly curtailing execution. The measured flexibility also varies with resource accounting, delay tolerance, and time resolution.

A. Temporal and Power Characteristics of AI Workloads

B. Peak Power Demand Shaving

  • During the 15:00–15:25 flexibility window, optimized profiles remain below original demand, showing peak shaving without directly curtailing workload execution.The reported comparison examines shifting under different delay rates.Source: Figure 4 · p. 4, B. Peak Power Demand Shaving · p. 4
  • At γ = 1.0, Case 2 flexibility is 32.47 kW versus 21.71 kW in Case 1, a 49.6% increase.Case 2 includes CPU, GPU, and memory power; Case 1 considers GPU power only.Source: B. Peak Power Demand Shaving · p. 4, Table I · p. 5
  • In Case 2, F rises from 10.76 kW at γ = 0 to 32.47 kW at γ = 1.0, about 2% to 7% of approximately 540 kW demand.Higher γ accompanies more shifted workloads and, at higher values, longer average shifting time.Source: B. Peak Power Demand Shaving · p. 4, Table I · p. 5
  • As temporal resolution becomes coarser, maximum achievable flexibility decreases because fewer workloads can be effectively shifted.Fig. 5 compares 5-, 15-, and 60-minute resolutions.Source: Figure 5 · p. 5, B. Peak Power Demand Shaving · p. 5

C. Sustained Power Demand Reduction

Sustained demand reduction involves a tradeoff: requiring greater power reduction shortens the achievable duration. The paper concludes that its framework quantifies both the magnitude and duration of AI data-center demand flexibility.

VI. CONCLUSIONS

  • The conclusion reports that greater workload delay tolerance increases short-term peak-shaving capability, while finer temporal resolutions preserve more scheduling flexibility.Source: §VI · p. 5

C. Sustained Power Demand Reduction

VI. CONCLUSIONS

  • The framework provides a practical means to quantify both the magnitude and duration of AI data-center demand flexibility for power system operation.Source: §VI · p. 5
Loading 2609.38020v1…