Predição de vazão em conexões TCP dedicadas ao streaming de vídeo de curta duração

Abstract

A recent forecasting study by Cisco showed that video transmission will account for 82% of all Internet traffic in 2022. This type of content is distributed in different forms and formats due to its non-linear consumption and audience fragmentation. In this context, the offering of short-duration videos, shorter than one minute, for example, has gained popularity, driven by video streaming platforms such as YouTube, Instagram, and TikTok. The distribution of this type of content presents particular challenges, including the limitation of end-to-end transfer rates imposed by congestion control mechanisms; whereas in longer sessions, it is possible to offer content with higher bit rates, that is, with better visual quality, since there is a certain stabilization of those congestion control mechanisms throughout the video session. The same reality is not observable in the transmission of short-duration videos. Industry efforts toward audience retention impose the need to always deliver content with the best possible quality. To this end, algorithms and heuristics employing bit-rate adaptation techniques are used for decision-making, whose cornerstone is the accurate prediction of throughput during an ongoing video session. The use of data from ongoing sessions by rate adaptation algorithms as the sole guiding source for decisions regarding the offered visual quality proves to be insufficient. In this work, a machine learning model is presented to predict throughput in connections dedicated to the transmission of short-duration videos using data collected from access histories. At its core, the model is characterized by a deep recurrent neural network that introduces artifacts to capture short-duration events by learning representative patterns of time-series signals. The results show that the proposed model achieves higher accuracy in scenarios where connections operate at an average speed above 2.88 Mbit/s, when compared to traditional moving average models and similar recurrent neural network models.