Source-linked AI summary
Full Scaling Automation for Sustainable Development of Green Data Centers
Shiyu Wang, Yinbo Sun, Xiaoming Shi, Shiyi Zhu, Lin-Tao Ma, James Zhang, Yifei Zheng, Jian Liu
TL;DR
Data-center emissions make efficient resource use important, while autoscaling faces challenges forecasting complex workloads and handling CPU-utilization uncertainty. FSA combines multi-scale workload representations, autoregressive forecasting, and task-conditioned Bayesian estimation to guide scaling. The authors report competitive real-world results and reduced energy consumption and carbon emissions after deployment.
Problem
Autoscaling must forecast complex, non-stationary workloads with varied periods and long temporal dependencies, a task the paper says existing methods do not handle well.
Source: §1 · pp. 1–2Method
FSA combines multi-scale time-series representations and representation-enhanced autoregressive forecasting with task-conditioned Bayesian estimation of workload and CPU-utilization relationships.
Source: §1 · p. 2Results
The authors report competitive real-world performance and reduced energy consumption and carbon emissions after deploying FSA in their data center.
Source: §6 · p. 7Takeaways & Limitations
The authors describe the deployment as an important step toward their company’s carbon-neutrality goal by 2030.
Source: §6 · p. 7The workload-per-Pod formulation assumes application load is evenly allocated across Pods, an assumption the authors say their data-center engineering platform ensures.
Source: §3.1 · p. 3
Abstract
from arXiv · showhide
The rapid rise in cloud computing has resulted in an alarming increase in data centers' carbon emissions, which now accounts for >3% of global greenhouse gas emissions, necessitating immediate steps to combat their mounting strain on the global climate. An important focus of this effort is to improve resource utilization in order to save electricity usage. Our proposed Full Scaling Automation (FSA) mechanism is an effective method of dynamically adapting resources to accommodate changing workloads in large-scale cloud computing clusters, enabling the clusters in data centers to maintain their desired CPU utilization target and thus improve energy efficiency. FSA harnesses the power of deep representation learning to accurately predict the future workload of each service and automatically stabilize the corresponding target CPU usage level, unlike the previous autoscaling methods, such as Autopilot or FIRM, that need to adjust computing resources with statistical models and expert knowledge. Our approach achieves significant performance improvement compared to the existing work in real-world datasets. We also deployed FSA on large-scale cloud computing clusters in industrial data centers, and according to the certification of the China Environmental United Certification Center (CEC), a reduction of 947 tons of carbon dioxide, equivalent to a saving of 1538,000 kWh of electricity, was achieved during the Double 11 shopping festival of 2022, marking a critical step for our company's strategic goal towards carbon neutrality by 2030.
1 Introduction
Data-center emissions and energy use make efficient resource allocation important, but existing autoscaling faces challenges in forecasting complex workloads, stabilizing CPU utilization under uncertainty, and accounting for carbon emissions. The paper proposes Full Scaling Automation (FSA), combining workload forecasting and uncertainty-aware scaling decisions to target CPU utilization.
1 Introduction
- Data centers account for 3.7% of global greenhouse gas emissions, motivating more efficient use of their computing resources.The introduction links emissions from large-scale clusters mainly to equipment power consumption and identifies chronically low CPU utilization as a source of energy waste.Source: §1 · p. 1
- Complex, non-stationary workloads with varied periods and long temporal dependencies challenge forecasting methods such as ARIMA and simple neural networks.The paper says most autoscaling studies focus on scaling decisions rather than accurate workload forecasting.Source: §1 · pp. 1–2
- Existing autoscaling work is described as optimizing resources and costs rather than carbon emissions, a sustainability challenge for green data centers.The paper also identifies stable CPU utilization and uncertainty-aware decisions as challenges for autoscaling.Source: §1 · p. 2
- FSA combines workload forecasting, CPU-utilization uncertainty modeling, and predictive horizontal autoscaling to target CPU utilization.Its forecasting framework learns multi-scale representations and uses a representation-enhanced deep autoregressive model; task-conditioned Bayesian regression estimates the workload–CPU relationship and its uncertainty.Source: §1 · p. 2
- FSA’s contributions include multi-scale workload forecasting, task-conditioned Bayesian estimation, and predictive horizontal autoscaling to stabilize CPU utilization and save energy.The proposed autoscaling mechanism is intended to facilitate sustainable data centers.Source: §1 · p. 2
2 Background And Related Work
The background reviews horizontal autoscaling and time-series forecasting, then introduces Bayesian regression as a way to estimate uncertainty in the workload–CPU relationship. It positions FSA against reactive scaling and forecasting approaches that do not address the paper’s stated needs.
2.1 Horizontal Autoscaling
- Prior horizontal autoscalers such as FIRM and Autopilot are described as reactive, lacking workload forecasts, and not modeling CPU-utilization uncertainty in scaling decisions.Horizontal autoscaling changes the number of assigned containers or Pods.Source: §2.1 · p. 2
2.2 Workload Forecasting
- Time-series forecasting work includes RNN- and Transformer-based models and representation-learning methods for temporal dependencies and periodicity; FSA uses representations to capture long-term dependencies.Figure 1 identifies the forecast module as multi-scale time-series representation plus a representation-enhanced deep autoregressive model.Source: §2.2 · pp. 2–3, Figure 1 · p. 3
2.3 Bayesian Method
- The paper introduces a task-conditioned hypernetwork into Bayesian regression to estimate the correlation between CPU utilization and workload.Bayesian methods are presented as a way to model uncertainty and improve robustness.Source: §2.3 · p. 3
3 Proposed Methodology
The methodology forecasts workload from compressed long-term time-series representations and recent observations, then uses task-conditioned workload estimates to choose Pod allocations around a target CPU utilization.
3 Proposed Methodology
- The proposed methodology comprises a workload forecast module and a decision module for selecting Pod allocations.The paper presents these modules as the overall framework.Source: §3 · p. 3, §3.3 · p. 5
3.1 Problem Definition
- Given historical workload observations, the allocation objective is to choose future Pod counts that keep an application near its target CPU utilization, assuming load balancing distributes workload evenly.The paper states that its data-center software and hardware platform ensures the even-allocation assumption.Source: §3.1 · p. 3
3.2 Workload Forecasting
- The method combines compressed long-term workload representations with short-term observations in a deep autoregressive forecasting model for future workload prediction.The model uses recurrent encoding, multi-head self-attention fusion, and recurrent autoregressive decoding.Source: §3.2 · p. 3, Representation-enhanced Deep… · p. 5
Multi-scale Time Series (TS) Representation
- The representation encoder learns time-domain trends and FFT-derived frequency patterns, then encodes one-minute series into daily, weekly, monthly, and quarterly historical vectors.Its contrastive learning uses views of the same timestamp as positives for time-domain learning.Source: Multi-scale Time Series (TS)… · p. 4
Representation-enhanced Deep Autoregressive Model
- The representation-enhanced autoregressive model uses long-term historical representations and near-window observations to predict future time-series values.The model fuses these inputs and performs multi-step autoregressive decoding.Source: Representation-enhanced Deep… · p. 5
3.3 Scaling Decision
- A task-conditioned Bayesian regression maps target CPU utilization and app identity to per-Pod workload, estimates uncertainty, and informs workload bounds; Pods are scaled every five minutes.Task embeddings condition the model's estimated parameter distribution for each app scenario.Source: §3.3 · p. 5
4 Experiment
Experiments on real-world industrial data evaluate workload forecasting, CPU-utilization and workload estimation, and pod-scaling. FSA uses the least resources among the tested autoscaling methods while maintaining target CPU utilization and stable SLO assurances.
4.1 Setup
- The evaluation covers forecasting, CPU–workload estimation, and pod-scaling using 10-minute data from 3,000 online microservices.The collected data includes workload, CPU utilization, and running Pod counts.Source: §4 · p. 5, §4.1 · pp. 5–6
Workload TS Forecast
- The workload forecaster achieves the best reported results on high-frequency real-world datasets, while representation-enhanced models outperform vanilla deep models by around 30%.The representations capture nested daily and weekly patterns, and self-attention fuses long-term representations with short-term observations.Source: Table 1 · p. 6, Workload TS Forecast · p. 6, Figure 2 · p. 6
CPU-utilization and Workload Estimation
- Task-conditioned Bayesian Neural Regression achieves the best reported results, with around 40% accuracy improvement while modeling CPU–workload uncertainty across scenarios.The model uses a globally unified model across scenarios, unlike the independent models used by several baselines.Source: Table 2 · p. 6, CPU-utilization and Workload Estimation · p. 6
Scaling Pods Decision
- FSA uses the least resources among the tested autoscaling methods while maintaining target CPU utilization and stable SLO assurances.Its decision range represents maximum benefits and minimum risk, with the choice between bounds guided by response-time tolerance.Source: Figure 3 · p. 7, Scaling Pods Decision · p. 7
5 Deployment
FSA was deployed on the company’s data-center server cluster, supporting more than 3,000 online application services. Across four Double 11 shopping festivals, certified reductions totaled 282, 335, 394, and 947 tons of carbon dioxide.
5 Deployment
- FSA was deployed on the company’s server cluster to support resource scheduling for more than 3,000 online application services.The reported deployment effectiveness and energy savings cover the past four years.Source: §5 · p. 7
- 282, 335, 394, and 947 tons of carbon dioxide were reduced across the four Double 11 festivals, equivalent to savings of 459,000, 545,000, 640,000, and 1,538,000 kWh, respectively.The reductions were certified by the China Environmental United Certification Center.Source: §5 · p. 7
6 Conclusion
FSA uses future-workload forecasts to adjust resources toward a desired target level in large-scale cloud computing clusters. Real-world tests showed competitiveness against state-of-the-art methods, and deployment reduced energy consumption and carbon emissions.
6 Conclusion
- FSA uses future-workload forecasts to automatically adjust resources toward a desired target level in large-scale cloud computing clusters.Source: §6 · p. 7
- Empirical tests on real-world data showed FSA was competitive with state-of-the-art methods, while deployment reduced energy consumption and carbon emissions.The authors describe the deployment as a step toward the company’s carbon-neutrality goal by 2030.Source: §6 · p. 7
Contribution Statement
The design, implementation, and deployment were carried out primarily by Shiyu Wang, Yinbo Sun, and Xiaoming Shi, whom the statement identifies as key contributors.
Contribution Statement
- Shiyu Wang, Yinbo Sun, and Xiaoming Shi carried out the design, implementation, and deployment primarily.Source: Contribution Statement · p. 7