Previsão do tamanho de enxames BitTorrent com redes neurais

Abstract

In peer-to-peer (P2P) networks it is important to understand the behavior of the entities that are part of the network or system. Only by fully understanding how the system really works is it possible to identify limitations and opportunities for improvement. Monitoring Bitorrent swarms is challenging as it requires a distributed system with a capacity that must be proportional to the size of the swarm. We propose to try to understand the swarm dynamics over time, based on the monitoring data itself. For this, we evaluated two prediction methods, the Temporal Graph Convolutional Neural Network (T-GCN) and the Autoregressive Integrated Moving Average Model (ARIMA), comparing which is the best choice to predict the size of the swarm over time. We found that the neural network can be a faster alternative to execute than ARIMA, although its result is a little worse.