Accurate Network Synchronization via Machine-Learning Based Path Asymmetry Estimation

  • Nada Alhashmi

Student thesis: Master's Thesis


Emerging technology trends in smart power grids, telecommunications, automated manufacturing, and high-frequency trading share the need for operating in a distributed and coordinated fashion, which requires synchronizing geographically-disparate assets with high accuracy. Rather than a costly dedicated infrastructure, a cost-effective approach is to utilize existing synchronization standards such as precision time protocol (PTP), where nodes can synchronize by exchanging messages containing timing information over general purpose networks. In this setting, the synchronization problem can be posed as an estimation problem from noisy observations, where the noise is introduced by several factors including synchronization hardware, propagation delays, and queuing-induced packet delay variations (PDVs). While existing approaches have improved synchronization performance subject to these variations, a significant limitation is their assumption that the message exchange paths are symmetric, i.e. of equal delay, which may never be satisfied in practice given the nondeterministic nature of network traffic. Moreover, joint estimation of synchronization and delay parameters is infeasible since it yields to an indeterminate system. As such, path asymmetry is currently a significant source of error affecting existing synchronization approaches. This research proposes a novel path asymmetry (PA) estimation scheme based on deep neural networks (DNNs). Specifically, a DNN configured as a Denoising Auto Encoder (DAE) is trained to reconstruct PA from real time measurements by separating the desired information from network-induced noise. This proposed solution can be used directly for path asymmetry estimation, or it can be integrated with different existing techniques to overcome their limitation and provide a collaboratively better synchronization system. Using simulation and industry-standard workloads, we demonstrate that the proposed system improves PA estimation and overall synchronization accuracy compared with existing approaches.
Date of AwardDec 2019
Original languageAmerican English
SupervisorNawaf Al Moosa (Supervisor)


  • Clock Synchronization
  • Precision Time Protocol (PTP)
  • Packet Delay Variation (PDV)
  • Deep Neural Network (DNN)
  • Convolutional Denoising Autoencoder (CDAE)
  • Path Asymmetry Estimation

Cite this