TL;DR
This paper introduces TUBO, a novel machine learning framework designed to overcome the challenges of bursty and dynamic network traffic patterns for more reliable traffic demand forecasting. By intelligently handling traffic spikes and dynamically selecting the best forecasting model, TUBO significantly improves accuracy and enables more effective proactive network management.
What Problem Does This Paper Address?
Accurately predicting network traffic is crucial for optimizing network operations, traffic engineering, and capacity planning. However, existing forecasting methods, even advanced deep learning models, struggle with the inherent characteristics of network traffic. These include:
- Bursty Traffic: Sudden, significant spikes in traffic volume that can overwhelm standard forecasting models and lead to inaccurate predictions.
- Dynamic Patterns: Traffic patterns that change over time due to various factors like seasonal changes, user behavior, or network events (e.g., link failures).
- Limited Adaptability: Many models are trained for specific patterns and fail to adapt when traffic characteristics shift, leading to a decline in performance.
This unreliability hinders the full potential of data-driven network management, particularly in proactive approaches like traffic engineering.
Methods
TUBO is a machine learning framework built around two core innovations:
- Burst Processing: This component specifically addresses traffic bursts. It first identifies and isolates outlier (burst) data points from the regular traffic data. This segregated burst data is then used to train a dedicated burst occurrence forecasting module. This module predicts whether a burst will occur in the next time slot. If a burst is predicted, downstream applications can rely on real-time measurements instead of potentially inaccurate forecasts. Non-burst data is used for training the main forecasting models.
- Model Selection: Recognizing that no single forecasting model is optimal for all traffic patterns, TUBO employs a model selection mechanism. It maintains a pool of diverse, pre-trained forecasting models (e.g., RNNs, Transformers). At inference time, for a given input traffic pattern, TUBO uses a post-training uncertainty measurement and calibration technique (specifically Monte Carlo Dropout) to estimate the predictive uncertainty of each model. The model that exhibits the lowest calibrated uncertainty, indicating it's most confident and likely to be accurate for that specific input, is selected for forecasting. This ensures the framework adapts to varying traffic patterns.
TUBO also incorporates flexible normalization techniques (Global, Individual, Rolling) to prepare data for model training, ensuring consistency and stability.
Datasets
The paper evaluates TUBO using three real-world network demand matrix (DM) datasets:
- Abilene: A dataset representing a network with 12 nodes and 30 links.
- GEANT: A dataset with 23 nodes and 74 links.
- CERNET: A dataset comprising 14 nodes and 32 links.
Additionally, they used TMGen to generate synthetic DM data for the Abilene, GEANT, and CERNET topologies to further validate their approach.
Key Insights and Ideas
- Bursts are Predictable (to an extent): While bursts are challenging, the paper shows they exhibit enough structure and correlation across OD pairs to be predicted with high accuracy (up to 94% burst occurrence rate). This predictability is key to managing their impact.
- No One-Size-Fits-All Model: The evaluation clearly demonstrates that different forecasting models excel on different datasets and even for different OD pairs within the same dataset. This strongly supports the need for dynamic model selection.
- Uncertainty is a Powerful Selection Criterion: Using uncertainty quantification (via Monte Carlo Dropout and distribution calibration) as a proxy for model reliability proves more effective than simply picking the model with the best average performance. It allows TUBO to adapt to unseen patterns and specific input characteristics.
- Proactive TE Benefits Greatly from Accurate Forecasting: The traffic engineering use case highlights that the gains from improved forecasting directly translate to significant improvements in network performance (throughput), especially when compared to reactive approaches.
Key Results
- Forecasting Accuracy: TUBO outperforms existing methods by up to 4x in terms of Mean Absolute Error (MAE) across the evaluated real-world datasets.
- Burst Occurrence Forecasting: Achieves up to 94% accuracy in predicting burst occurrences.
- Traffic Engineering (TE) Improvement:
- Proactive TE with TUBO improves aggregate throughput by 9x compared to reactive TE (SMORE).
- It also shows a 3x improvement over proactive TE using the best prior DM forecasting method (ConvLSTM).
- Model Selection Impact: TUBO's model selection process can improve average OD pair MAE by up to 2x and worst-case MAE by up to 4x.
Limitations and Future Work
- Scalability for Extremely Large Networks: While the paper demonstrates TUBO's scalability for up to 20,000 nodes within a minute granularity, extremely large-scale deployments might still present computational challenges.
- Concept Drift: The paper acknowledges that handling sudden distribution shifts (concept drifts) in traffic data is an ongoing challenge. TUBO's framework is designed to be orthogonal to this, meaning it can benefit from improved data preprocessing or retraining strategies for concept drift.
- Future Work: The authors intend to open-source the TUBO framework. Further research could explore more advanced methods for uncertainty quantification and calibration, as well as investigate the integration of TUBO with other network management tasks beyond TE.
Summary updated