Source-linked AI summary
EsDNN: Deep Neural Network based Multivariate Workload Prediction Approach in Cloud Environment
Minxian Xu, Chenghao Song, Huaming Wu, Sukhpal Singh Gill, Kejiang Ye, Chengzhong Xu
TL;DR
High variance and dimensionality make cloud workload prediction difficult, complicating efficient resource provisioning. esDNN converts multivariate time series into supervised data and applies a revised GRU; experiments on Alibaba and Google traces report better accuracy than state-of-the-art approaches.
Problem
High variance and dimensionality make cloud workloads difficult to predict accurately, although forecasts can support more efficient resource allocation.
Source: §1 · p. 2Method
esDNN uses a sliding window to convert multivariate time series into supervised data and applies a revised GRU-based prediction algorithm.
Source: §1.1 · p. 3, §4.1 · p. 9Results
Experiments on realistic Alibaba and Google traces report better prediction accuracy than state-of-the-art approaches.
Source: §6 · p. 22Takeaways & Limitations
In Google auto-scaling simulations at minute 400, the baseline required 18,371 machines while esDNN used 5,161.
Source: §5.4 · p. 22
Abstract
from arXiv · showhide
Cloud computing has been regarded as a successful paradigm for IT industry by providing benefits for both service providers and customers. In spite of the advantages, cloud computing also suffers from distinct challenges, and one of them is the inefficient resource provisioning for dynamic workloads. Accurate workload predictions for cloud computing can support efficient resource provisioning and avoid resource wastage. However, due to the high-dimensional and high-variable features of cloud workloads, it is difficult to predict the workloads effectively and accurately. The current dominant work for cloud workload prediction is based on regression approaches or recurrent neural networks, which fail to capture the long-term variance of workloads. To address the challenges and overcome the limitations of existing works, we proposed an efficient supervised learning-based Deep Neural Network (esDNN}) approach for cloud workload prediction. Firstly, we utilize a sliding window to convert the multivariate data into supervised learning time series that allow deep learning for processing. Then we apply a revised Gated Recurrent Unit (GRU) to achieve accurate prediction. To show the effectiveness of esDNN, we also conduct comprehensive experiments based on realistic traces derived from Alibaba and Google cloud data centers. The experimental results demonstrate that esDNN can accurately and efficiently predict cloud workloads. Compared with the state-of-the-art baselines, esDNN can reduce the mean square errors significantly, e.g. 15% than the approach using GRU only. We also apply esDNN for machines auto-scaling, which illustrates that esDNN can reduce the number of active hosts efficiently, thus the costs of service providers can be optimized.
1 INTRODUCTION
Cloud workload prediction matters because workload variability and dimensionality make accurate forecasting difficult, while forecasts can support advance resource allocation. The paper combines multivariate sliding-window conversion with esDNN and evaluates prediction and auto-scaling.
1 INTRODUCTION
- Mismatched provisioning can waste resources during low workloads or degrade performance during sharp increases, motivating accurate forecasts for advance allocation and de-allocation.Forecast-guided planning is presented as a way to use resources more efficiently.Source: §1 · p. 2
- Workload variance and high-dimensional traces make cloud workloads difficult to forecast; Alibaba resource utilization ranges from 5% to 80%.Traces also contain information about different machine specifications that must be assessed for model training.Source: §1 · p. 2
- Regression methods and shallow neural networks work best on workloads with obvious patterns but struggle to capture complicated workload correlations in large public clouds.The introduction therefore motivates using more complicated neural networks to exploit those correlations.Source: §1 · p. 2
1.1 Motivation and Our Contributions
- The paper proposes S-MTF to convert multivariate time series into supervised learning data and esDNN, a revised GRU-based predictor evaluated on Alibaba and Google traces.The contributions also include evaluating esDNN for auto-scaling.Source: §1.1 · p. 3
1.2 Article Organization
- The article proceeds from related work and the system model to the DNN algorithm, experiments, and conclusions with future directions.Source: §1.2 · p. 3
2 RELATED WORK
Prior cloud workload prediction research includes regression and learning-based methods, but the paper identifies difficulties handling variable, high-dimensional workloads and long-term dependencies. It positions its GRU and sliding-window approach as addressing these challenges while retaining workload information.
2.1 Regression-based Approaches for Cloud Workload Prediction
- Regression approaches can predict workloads with obvious patterns, but high-variance cloud workloads make their parameter correlations difficult to represent.Prior evaluations often used high-performance computing, small-scale, or synthetic workloads with lower variance than cloud workloads.Source: §2.1 · p. 4
2.2 Learning-based Approaches for Cloud Workload Prediction
- Learning-based work spans neural networks, deep learning, recurrent models, attention, and online adaptation, including GRU-based cloud prediction paired with auto-encoder compression.The paper notes that compression can address dimensionality but may undermine accuracy by not using the full workload data.Source: §2.2 · pp. 4–5
2.3 Critical Analysis
- The paper combines GRU-based long-term dependency modeling with a multivariate sliding window that converts time series into supervised data while retaining information and predicted-data correlations.It validates the approach using realistic Google and Alibaba traces and multiple metrics.Source: §2.3 · p. 6
- The paper's critical analysis frames its contribution as addressing high-dimensional and multivariate prediction through GRU and sliding-window processing, with related work compared in Table 1.Source: §2.3 · pp. 5–6, Table 1 · p. 6
3 SYSTEM MODEL
The system model predicts future cloud workloads so providers can optimize resource use and costs, including through auto-scaling. It preprocesses and reframes multivariate traces, trains a GRU-derived model, then uses predictions to adjust machine provisioning.
3 SYSTEM MODEL
- The model targets accurate, efficient workload prediction to support resource optimization and potentially reduce costs by decreasing active hosts through auto-scaling.It uses CPU usage as the standard for measuring prediction results and models multivariate time series because workload signals can be correlated.Source: §3 · p. 6
- Preprocessing groups data by timestamp, computes parameter averages, and applies MinMaxScaler to normalize datasets before modeling.The predicted resource utilization ranges from 0.0 to 1.0, and missing values are filled using the previous time slot.Source: §3 · pp. 7–8
- The supervised-learning conversion reframes normalized multivariate time series into samples X and corresponding labels Y, using training and validation sets.Source: §3 · p. 8
- The model-construction step builds a GRU-derived network from preprocessed data; trained predictions can then guide dynamic machine additions or removals.Alibaba predictions are usually 10 seconds apart, while Google predictions are 5 minutes apart.Source: §3 · p. 8
4 ESDNN: EFFICIENT SUPERVISE LEARNING-BASED DEEP NEURAL NETWORK
esDNN converts multivariate workload time series into supervised sequences, then applies a GRU-based deep-learning network to predict future workloads. Its S-MTF transform retains prior-time data, while the network combines convolutional processing with recurrent modeling and uses Swish activation.
4 ESDNN: EFFICIENT SUPERVISE LEARNING-BASED DEEP NEURAL NETWORK
- esDNN converts multivariate workload time series into supervised-learning sequences before applying a GRU-based network for future workload prediction.The approach first applies a sliding window to the original dataset, then trains the network on the converted data.Source: §4 · p. 8
4.1 Sliding Window for Multivariate Time Series Forecasting
- S-MTF forms supervised samples and labels from multivariate data, using prior and current observations as samples and the next observation as the label.The examples include one-step multivariate forecasting, with memory utilization selected as the label.Source: §4.1 · p. 9
- S-MTF retains all values from the previous moment for optional selection and can include future labels for multi-step forecasting.The authors state that its processing remains linearly related to time and that it transforms time-related data into sample and label datasets.Source: §4.1 · p. 10
4.2 esDNN Algorithm
- The network adds a 1D convolutional layer to extract local time-series features before the GRU-based layer models sequence relationships.Causal zero-padding is used to preserve boundary information and support prediction of early time steps.Source: §4.2 · p. 12
- esDNN uses Swish instead of ReLU in its deep network, combining ReLU and sigmoid through self-gating with a shared value.The authors describe self-gating as requiring a scalar input while allowing replacement of single-scalar activation functions without changing hidden capacity or parameter count.Source: §4.2 · p. 13
5 PERFORMANCE EVALUATION
The evaluation tests esDNN on Alibaba and Google traces, compares it with unsupervised and RNN-based approaches, and simulates its use for machine auto-scaling.
5.1 Datasets and Environment Configuration
- The experiments use Alibaba traces from about 4,000 machines over eight days and Google traces from 37,747 machines over 29 days, with five-minute intervals and dataset-specific input features.Alibaba training uses the first 40,000 rows; Google uses about 72 hours, with the first 49 hours for training. Figures 5 and 6 show daily and minute-level CPU fluctuations.Source: §5.1 · pp. 14–15, Figure 5 · p. 15, Figure 6 · p. 15
5.2 Comparison with Unsupervised Learning-based Approach
- Within 10 minutes, Autoencoder is less accurate than esDNN except at the beginning of the observed period.The authors suggest Autoencoder’s reconstruction-based training may make complex models prone to overfitting.Source: §5.2 · p. 16, Figure 7 · p. 16
5.3 Comparison with Neural Network-based Approaches
- Across RNN baselines, esDNN’s MSE is lower overall than Bi-LSTM and GRU on Alibaba, and it is more accurate and stable than GRU in the long run.On Google, esDNN and GRU have close MSE; esDNN is reported as more stable.Source: §5.3 · pp. 18–19, Figure 9 · p. 18, Figure 10 · p. 18
- esDNN achieves the lowest MSE for longer prediction lengths and the lowest RMSE on Google at longer lengths; its MSE also falls by one order of magnitude versus.The reported reduction is from around 7 × 10^-2 to 7 × 10^-3.Source: §5.3 · pp. 19–20, Table 6 · p. 20, Table 7 · p. 20
5.4 Applying esDNN for Machines Auto-scaling with Simulations
- In the Google simulation, esDNN uses 5,161 machines versus 18,371 for the baseline at minute 400, while Alibaba’s ratio indicates 40%–80% of the original machines remain active.The Google ratio ranges from 0.25 to 1; the Alibaba ratio fluctuates from 0.3 to 0.6.Source: §5.4 · pp. 21–22, Figure 14 · p. 21
6 CONCLUSIONS AND FUTURE WORK
On realistic Google and Alibaba traces, esDNN improves prediction accuracy over state-of-the-art approaches and optimizes resource use through auto-scaling. Future work targets container integration, edge and mobile-aware scenarios, and MAPE-based automation.
6 CONCLUSIONS AND FUTURE WORK
- On realistic Google and Alibaba traces, esDNN achieves better prediction accuracy than state-of-the-art approaches.Source: §6 · p. 22
- Applying esDNN to auto-scaling on realistic traces significantly optimizes data-center resource use, saving operational costs.Source: §6 · p. 22
- A sliding window converts multivariate data into supervised time series, which a revised GRU-based model uses to predict high-variance workloads.Source: §6 · p. 22
- Future work proposes integrating esDNN into a container-based prototype such as Kubernetes and extending it to edge computing with offloading.Source: §6 · p. 22
- The authors also plan to study location- and mobility-aware workload prediction and automate esDNN using the MAPE model.Source: §6 · p. 22