SM-FPLF: Link-State Prediction for Software-Defined DCN Power Optimization

Mohammed Nsaif, Gergely Kovasznai, Ali Malik, Ruairi De Frein

Research output: Contribution to journalArticlepeer-review

Abstract

Efficient monitoring systems that optimize resource allocation, reduce energy usage through machine learning and flow aggregation routing techniques, are needed due to the escalating power consumption of data center networks, which, as has been recently reported, account for up to eight percent of global energy consumption, posing environmental operational concerns. We propose a software-defined data-center monitoring algorithm that reduces power consumption by: 1) using a GPU implementation of a Stacked Long Short-Term Memory Recurrent Neural Network (RNN) model for link utilization prediction, thus reducing monitoring overhead; and 2) utilizing a flow aggregation routing algorithm with feedback from online, OpenFlow-powered monitoring and machine learning modules. This combined approach results in a new algorithm called SMart-Fill Prefer Path First (SM-FPLF). In the context of SM-FPLF, the objective of this paper is to compare the: 1) training and validation loss curves for various models; 2) to evaluate the prediction accuracy of learning approaches for a range of prediction horizons; 3) to assess the time-cost and accuracy for different models, with a specific focus on the GuSLSTM and GuGRU models; 4) to analyze OpenFlow traffic with and without using the preferred prediction algorithm, the GuSLSTM model, assessing the accumulated power consumption per OpenFlow channel in the data-centre when SM-FPLF is applied. Our findings indicate that the GuSLSTM outperforms rival algorithms in terms of link utilization prediction accuracy over varying input sequence lengths. This accuracy is achieved whilst satisfying the SDN domain-specific requirement of a small computation time in a real-time implementation. Embedding a GuSLSTM in the SM-FPLF algorithm offers a power saving of 372 watts per OpenFlow channel, which is achieved in part due to a 13.7% CPU usage reduction in controllers and switches. These findings provide a valuable perspective into the performance and suitability of RNNs for real-time implementation as part of SDN solutions. They also shed light on their practical implications and benefits of using link utilization prediction in SDN management and power consumption optimization solutions.

Original languageEnglish
Pages (from-to)79496-79518
Number of pages23
JournalIEEE Access
Volume12
DOIs
Publication statusPublished - 2024

Keywords

  • Data center networks
  • machine learning
  • monitoring
  • OpenFlow
  • overhead
  • power consumption
  • prediction
  • software-defined networks

Fingerprint

Dive into the research topics of 'SM-FPLF: Link-State Prediction for Software-Defined DCN Power Optimization'. Together they form a unique fingerprint.

Cite this