Source-linked AI summary
A Two-Stage Forecasting System for CPU Workload Prediction in Private Clouds
Ashir Javeed, Anton Borg, Håkan Grahn, Lars Lundberg, Dhyey Patel, Sogand Shirinbab
TL;DR
Cloud resource forecasting must support QoS and efficient provisioning, yet direct CPU-trace methods can overlook the relationship between customer demand and CPU consumption. This study uses a two-stage XGBoost system that forecasts TPS before CPU workload, with expanding-window retraining for evolving workloads. Evaluated on ten private-cloud applications, it achieved a median SMAPE of 5.9%, sub-264 ms integrated latency, and stable recursive errors across 60 steps.
Problem
Direct CPU forecasting from historical resource traces can overlook customer service demand, although that demand is linked to downstream CPU consumption.
Method
The system uses cascaded XGBoost models to forecast TPS first and estimate CPU workload second, with walk-forward expanding-window retraining for concept drift.
Results
A median SMAPE of 5.9% was achieved for the complete forecasting model, with integrated-pipeline latency below 264 ms and stable recursive errors across 60 steps.
Takeaways & Limitations
The evaluated system supports proactive CPU resource prediction with computational efficiency across diverse workloads in ten private-cloud applications.
Takeaways & Limitations
Evaluation used TPS-driven data from a private cloud under relatively stable conditions, focused only on CPU, and had limited dataset size and workload diversity.
Abstract
from arXiv · showhide
Accurate cloud resource forecasting is essential for proactive resource provisioning, maintaining Quality of Service (QoS), and reducing operational costs in dynamic cloud environments. The existing forecasting approaches predominantly estimate future CPU workload directly from historical resource traces, which often overlook the relationship between customer service demand and subsequent resource consumption. This study proposes a two-stage integrated forecasting model that explicitly models this dependency by first forecasting customer service requests, expressed as Transactions Per Second (TPS), and subsequently estimating future CPU workload from the TPS forecast. Both the forecasting component and resource prediction component employed the XGBoost model within a cascaded learning architecture, complemented by adaptive online retraining using an expanding-window strategy to address concept drift in continuously evolving cloud workloads. The proposed work was evaluated using real-world traces collected from a private cloud environment comprising ten applications. Experimental results demonstrate robust forecasting performance by achieving Symmetric Mean Absolute Percentage Error (SMAPE) below $7\%$ for most applications, with the best-performing application achieving an MAE of $0.7372$, RMSE of $1.1866$, SMAPE of $3.57\%$, and an R2 of $0.9185$. Horizon-wise drift analysis confirmed stable recursive forecasting behavior with controlled error accumulation across a 60-step prediction horizon. Compared with the conventional direct CPU forecasting method, the proposed two-stage integrated model gives improved forecasting robustness, computational efficiency, and interpretability, making it well-suited for proactive resource management and intelligent auto-scaling in cloud computing environments.
1 Introduction
Cloud resource forecasting supports QoS and efficient provisioning, but direct CPU forecasting can overlook customer-demand dynamics. The study proposes a two-stage, TPS-driven system with adaptive retraining and XGBoost-based learning for robust, efficient prediction.
- Over-provisioning wastes energy, whereas under-provisioning can harm QoS, making efficient cloud-resource allocation necessary.
- 5% to 80% variation in Alibaba cloud workloads illustrates the difficulty of forecasting rapidly changing resource demand.
- Existing methods often forecast CPU utilization directly from historical traces, while customer-request effects on downstream CPU demand remain insufficiently modeled.
- The proposed system forecasts customer service requests as TPS before estimating downstream CPU utilization in two sequential learning stages.
- Adaptive online retraining and Rolling Max smoothing were added to improve stability under highly dynamic workloads.
- XGBoost-based learning provides computationally efficient forecasting intended for lightweight streaming deployment.
2 Related Work
Cloud resource-forecasting research spans statistical, machine-learning, deep-learning, and hybrid approaches, but existing methods face challenges in modeling nonlinear workloads, computational cost, adaptability, and the link between service demand and CPU consumption.
- Approaches: Statistical, conventional machine-learning, and newer techniques have been applied to cloud resource forecasting for efficient resource management.Statistical models identify trends and patterns, while machine-learning approaches target forecasting accuracy.
- Approaches: Machine-learning models can learn nonlinear workload patterns but may require feature engineering, hyperparameter tuning, and substantial computation.These requirements can increase the cost of applying machine learning to cloud forecasting.
- Representative studies: Prior studies explored optimized XGBoost allocation, autoregressive workload forecasting, recurrent neural networks, and ensemble-based burst-aware autoscaling.These methods targeted forecasting accuracy, resource utilization, Quality of Service, or operational costs in cloud environments.
- Representative studies: Bayesian neural networks, probabilistic LSTM models, and transfer learning were used to forecast multiple cloud resources across Google and Alibaba datasets.The reported evaluation emphasized forecasting accuracy and service-level performance across different cloud data centres.
- Open challenges: Most current methods directly forecast CPU consumption from historical resource traces, potentially missing how inbound workload intensity and customer service demand relate to downstream CPU use.The stated concern is that direct forecasting may inadequately capture this relationship.
- Open challenges: Existing research also leaves real-time adaptability insufficiently addressed while statistical models struggle with nonlinear patterns and machine-learning approaches demand substantial resources and training data.These challenges accompany the field’s emphasis on improving prediction accuracy.
3 Methodology
The methodology uses a two-stage forecasting pipeline that first models TPS and then estimates CPU utilization from forecasted TPS. It combines walk-forward validation, periodic expanding-window retraining, and private-cloud workload data to support dynamic forecasting evaluation.
- Proposed system: The system separates TPS forecasting from downstream CPU resource estimation in two sequential learning stages.Forecasted customer requests serve as the intermediate input for CPU prediction.
- Stage 1: TPS Forecasting Module: The TPS forecasting module recursively predicts future request trajectories over the selected horizon.A multi-output model is trained from historical TPS windows, and each forecast is used in subsequent recursive steps.
- Stage 2: CPU Resource Estimation Module: The CPU estimation module directly consumes forecasted TPS features to produce future CPU utilization estimates.The CPU target uses a rolling maximum window to capture near-future peak demand while reducing sensitivity to short-term fluctuations.
- Adaptive retraining: Retraining occurs every r=1000 observations with an expanding window to balance adaptation speed, accuracy, and computational overhead.The strategy is intended to adapt to evolving workload distributions and mitigate concept drift in streaming environments.
- Dataset Collection and Environment: The dataset contains private-cloud CPU and service-request measurements collected at 1-minute intervals across selected applications.The original 1,440 samples were replicated and mirrored to produce approximately 8,514 samples, while some attribute definitions and units were withheld for privacy.
- Validation and Evaluation Metrics: Walk-forward validation trains on historical data and evaluates subsequent unseen observations while preserving temporal order.This procedure simulates deployment conditions and avoids information leakage from future data; evaluation includes RMSE, MAE, SMAPE, asymmetric error measures, and R2.
4 Experimental Results and Analysis
Across ten cloud applications, the two-stage model maintained low errors, strong explanatory power, and sub-second prediction runtime. Horizon-wise analysis showed controlled recursive error accumulation, while performance varied with workload stability and scale.
- MAE remained below 1.5 across eight applications, while B1 exhibited larger error because of higher workload variability.
- SMAPE stayed below 7% for applications A, B1, B2, B3, C, and D1, whereas D2 and D4 reached 25.48% and 45.73%.
- 0.9185 was application A’s highest R2 value, indicating that the model explained over 91% of workload variance.
- 178.50 ms to 263.70 ms was the runtime range, with an approximate 201 ms mean across applications.
- Application A achieved MAE 0.7372, RMSE 1.1866, SMAPE 3.57%, and R2 0.9185 in the graphical evaluation.
- Across the 60-step horizon, TPS drift curves rose smoothly without abrupt divergence or exponential error growth.
5 Discussion
The discussion attributes the system’s performance to hierarchical TPS-to-CPU modeling and evaluates the effects of forecasting uncertainty and runtime. It also identifies limitations in scope, workload conditions, dataset diversity, and expanding-window retraining cost.
- The two-stage system first forecasts TPS and then estimates CPU demand, explicitly modeling workload evolution before resource prediction.
- Forecasted TPS introduces uncertainty into downstream CPU prediction, producing lower accuracy than the setting using original TPS values.
- About 200 ms is required to construct a full 60-step forecasting horizon, supporting dynamic allocation and real-time monitoring without appreciable scheduling delays.
- The approach combines interpretability, long-horizon drift assessment, forecasting robustness, computational efficiency, and prediction accuracy.
- The evaluation covered only TPS-driven CPU prediction from a private cloud, excluding memory, storage, network bandwidth, and energy consumption.
- The expanding-window retraining strategy incurs increasing computational overhead as the training dataset grows.
6 Conclusion
The study develops a two-stage CPU forecasting system that predicts customer demand before estimating downstream CPU workload, using XGBoost and adaptive expanding-window retraining. Evaluation on ten private-cloud applications shows accurate, low-latency forecasting and stable recursive behavior across a 60-step horizon, with benefits for proactive resource management.
- Conclusion: The model first forecasts Transactions Per Second (TPS) and then estimates downstream CPU workload using XGBoost in both stages.Adaptive online retraining with an expanding window addresses changing workload patterns while keeping computational overhead low.
- Conclusion: 5.9% median SMAPE, R2 between 0.9145 −0.9903, and forecasting latency below 264 ms demonstrate accurate and efficient CPU forecasting across ten cloud applications.The CPU estimation latency was less than 1 ms, supporting real-time deployment in cloud auto-scaling systems.
- Conclusion: Stable recursive errors across a 60-step horizon for 3 out of 5 features support the feasibility of the cascaded forecasting architecture.The horizon-wise drift study examined error accumulation over successive prediction steps.
- Conclusion: The approach enables earlier anticipation of workload changes, supporting proactive provisioning, fewer SLA violations, improved utilization, and lower operating costs.XGBoost is presented as a computationally efficient alternative to deep learning-based forecasting models.
Authorship contribution
The paper assigns conceptualization, methodology, validation, supervision, resources, data curation, and writing responsibilities across six authors.
- Authorship contribution: Ashir Javeed handled conceptualization, methodology, validation, and the original draft.
- Authorship contribution: Anton Borg, H˚akan Grahn, Lars Lundberg, Dhyey Patel, and Sogand Shirinbab contributed to methodology, supervision, resources, data curation, validation, or manuscript review.
- Authorship contribution: Writing responsibilities included original drafting by Ashir Javeed and review and editing by Anton Borg, H˚akan Grahn, Lars Lundberg, Dhyey Patel, and Sogand Shirinbab.
Data availability
The study reports that its data is confidential.
- Data availability: The data used in the study is confidential.
- Data availability: The reported data-availability statement identifies confidentiality as the study's data constraint.
- Data availability: No additional data-access condition is stated in the supplied passage.