Forecasting short-term data center network traffic load with convolutional neural networks.

Area of Science:

• Computer Science
• Network Engineering
• Machine Learning

Background:

• Efficient data center resource management is critical for content service providers, with network traffic projected to increase significantly.
• Accurate forecasting of network traffic is essential for dynamic infrastructure adjustments and optimizing virtual machine activity.

Purpose of the Study:

• To propose and evaluate a convolutional neural network (CNN) model for short-term data center network traffic forecasting.
• To address the limitations of traditional time-series analysis methods like ARIMA in capturing chaotic network traffic patterns.

Main Methods:

• Utilizing convolutional neural networks (CNNs) to exploit non-linear regularities in network traffic data.
• Employing a multiresolution input strategy within the CNN architecture, distributing data across separate channels in the first convolutional layer.
• Validating the model using a 5-month dataset of one-second resolution traffic data from an Internet Service Provider's core network.

Main Results:

• The CNN model demonstrated significant improvements in forecasting accuracy compared to traditional methods.
• The CNN approach outperformed ARIMA, especially at forecasting granularities above 16 seconds.
• The multiresolution input enhanced the CNN's ability to capture complex traffic dynamics.

Conclusions:

• CNNs offer a superior approach for short-term network traffic forecasting in data centers.
• The proposed CNN model with multiresolution input provides a robust solution for efficient data center resource management.
• Accurate traffic prediction using CNNs can lead to optimized infrastructure shaping and improved service delivery.