Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Congestion or No Congestion: Packet Loss Identification and Prediction Using Machine Learning

Inayat Ali, Seungwoo Hong, Taesik Cheung

TL;DR

The paper addresses TCP throughput degradation caused by non-congestive wireless losses by differentiating congestive from non-congestive packet losses at the end host using data-driven features. It evaluates five classifiers (RF, KNN, GB, LR, DT) on ns-3 generated data with an 80/20 split and optimizes for recall due to class imbalance, aiming to enable TCP to avoid unnecessary cwnd reductions for non-congestive losses. RF and KNN emerge as the strongest performers, achieving high recall and balanced $F1$ scores, with jitter, RTT, and especially $cWnd$ identified as critical features via ablation studies. The work suggests that integrating such end-host loss-type identification can significantly boost wireless throughput in hybrid networks, guiding future deployment in dynamic environments.

Abstract

Packet losses in the network significantly impact network performance. Most TCP variants reduce the transmission rate when detecting packet losses, assuming network congestion, resulting in lower throughput and affecting bandwidth-intensive applications like immersive applications. However, not all packet losses are due to congestion; some occur due to wireless link issues, which we refer to as non-congestive packet losses. In today's hybrid Internet, packets of a single flow may traverse wired and wireless segments of a network to reach their destination. TCP should not react to non-congestive packet losses the same way as it does to congestive losses. However, TCP currently can not differentiate between these types of packet losses and lowers its transmission rate irrespective of packet loss type, resulting in lower throughput for wireless clients. To address this challenge, we use machine learning techniques to distinguish between these types of packet losses at end hosts, utilizing easily available features at the host. Our results demonstrate that Random Forest and K-Nearest Neighbor classifiers perform better in predicting the type of packet loss, offering a promising solution to enhance network performance.

Congestion or No Congestion: Packet Loss Identification and Prediction Using Machine Learning

TL;DR

The paper addresses TCP throughput degradation caused by non-congestive wireless losses by differentiating congestive from non-congestive packet losses at the end host using data-driven features. It evaluates five classifiers (RF, KNN, GB, LR, DT) on ns-3 generated data with an 80/20 split and optimizes for recall due to class imbalance, aiming to enable TCP to avoid unnecessary cwnd reductions for non-congestive losses. RF and KNN emerge as the strongest performers, achieving high recall and balanced scores, with jitter, RTT, and especially identified as critical features via ablation studies. The work suggests that integrating such end-host loss-type identification can significantly boost wireless throughput in hybrid networks, guiding future deployment in dynamic environments.

Abstract

Packet losses in the network significantly impact network performance. Most TCP variants reduce the transmission rate when detecting packet losses, assuming network congestion, resulting in lower throughput and affecting bandwidth-intensive applications like immersive applications. However, not all packet losses are due to congestion; some occur due to wireless link issues, which we refer to as non-congestive packet losses. In today's hybrid Internet, packets of a single flow may traverse wired and wireless segments of a network to reach their destination. TCP should not react to non-congestive packet losses the same way as it does to congestive losses. However, TCP currently can not differentiate between these types of packet losses and lowers its transmission rate irrespective of packet loss type, resulting in lower throughput for wireless clients. To address this challenge, we use machine learning techniques to distinguish between these types of packet losses at end hosts, utilizing easily available features at the host. Our results demonstrate that Random Forest and K-Nearest Neighbor classifiers perform better in predicting the type of packet loss, offering a promising solution to enhance network performance.
Paper Structure (12 sections, 1 figure, 3 tables)

This paper contains 12 sections, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Statement
  4. Dataset and Loss characterization
  5. Packet loss and feature selection
  6. Proposed Methodology
  7. Machine learning models
  8. Results
  9. Ablation study
  10. Single Feature Removal
  11. Multiple Feature Removal
  12. Conclusion

Figures (1)

  • Figure 1: Comaprison of stationary/mobile wireless client with wired client