Recent years have witnessed a remarkable upsurge in the global proliferation of Real-Time Communications (RTC) applications, a trend propelled by the flourishing advancement of network technologies and further amplified by the COVID-19 pandemic. Within this context, there is a burgeoning interest in the innovation of sophisticated and intelligent network infrastructures and technologies. Positioned as a promising candidate for this purpose, real-time throughput prediction emerges as a key enabler to foster network observability and offer proactive functions, upholding advanced system management, including but not limited to, bandwidth allocation and adaptive streaming. Nonetheless, existing methodologies struggle with predicting extreme conditions of throughput, notably peaks, valleys, and abrupt changes, that are critical in RTC traffic. To surmount these obstacles, we introduce DeX, a Deep Learning (DL)-based framework, designed to predict short-term throughput, with a dexterous proficiency and dedicated focus on navigating the complexities of traffic eXtremes. In particular, DeX leverages solely packet-level information as features and is composed of three integral components: a packet selection module that opts for an optimal subset of input features, a feature extraction block that partially incorporates the Transformer architecture, and a multi-task learning pipeline that improves the proficiency in handling traffic extremes. Moreover, our work is anchored in extensive traffic traces garnered during actual video-teleconferencing calls, and we formulate a time-series regression problem, rigorously evaluating a spectrum of technologies ranging from an adaptive filter to diverse Machine Learning (ML) and DL approaches. Initially, we aim at predicting throughput within 500-ms time windows using historical 1024 packets out of 2048, and consequently, our methodology exhibits exceptional efficacy, especially in forecasting traffic extremities. Conclusively, we conduct a series of ablation experiments and thorough analyses to showcase the enhanced performance of various scenarios, further validating the effectiveness and robustness of DeX.

DeX: Deep learning-based throughput prediction for real-time communications with emphasis on traffic eXtremes / Song, Tailai; Garza, Paolo; Meo, Michela; Munafo, Maurizio Matteo. - In: COMPUTER NETWORKS. - ISSN 1389-1286. - ELETTRONICO. - 249:(2024), p. 110507. [10.1016/j.comnet.2024.110507]

DeX: Deep learning-based throughput prediction for real-time communications with emphasis on traffic eXtremes

Song, Tailai;Garza, Paolo;Meo, Michela;Munafo, Maurizio Matteo
2024

Abstract

Recent years have witnessed a remarkable upsurge in the global proliferation of Real-Time Communications (RTC) applications, a trend propelled by the flourishing advancement of network technologies and further amplified by the COVID-19 pandemic. Within this context, there is a burgeoning interest in the innovation of sophisticated and intelligent network infrastructures and technologies. Positioned as a promising candidate for this purpose, real-time throughput prediction emerges as a key enabler to foster network observability and offer proactive functions, upholding advanced system management, including but not limited to, bandwidth allocation and adaptive streaming. Nonetheless, existing methodologies struggle with predicting extreme conditions of throughput, notably peaks, valleys, and abrupt changes, that are critical in RTC traffic. To surmount these obstacles, we introduce DeX, a Deep Learning (DL)-based framework, designed to predict short-term throughput, with a dexterous proficiency and dedicated focus on navigating the complexities of traffic eXtremes. In particular, DeX leverages solely packet-level information as features and is composed of three integral components: a packet selection module that opts for an optimal subset of input features, a feature extraction block that partially incorporates the Transformer architecture, and a multi-task learning pipeline that improves the proficiency in handling traffic extremes. Moreover, our work is anchored in extensive traffic traces garnered during actual video-teleconferencing calls, and we formulate a time-series regression problem, rigorously evaluating a spectrum of technologies ranging from an adaptive filter to diverse Machine Learning (ML) and DL approaches. Initially, we aim at predicting throughput within 500-ms time windows using historical 1024 packets out of 2048, and consequently, our methodology exhibits exceptional efficacy, especially in forecasting traffic extremities. Conclusively, we conduct a series of ablation experiments and thorough analyses to showcase the enhanced performance of various scenarios, further validating the effectiveness and robustness of DeX.
File in questo prodotto:
File Dimensione Formato  
1-s2.0-S1389128624003396-main.pdf

accesso aperto

Tipologia: 2a Post-print versione editoriale / Version of Record
Licenza: Creative commons
Dimensione 4.5 MB
Formato Adobe PDF
4.5 MB Adobe PDF Visualizza/Apri
Throughput_prediction_ComNet_final.pdf

accesso aperto

Tipologia: 2. Post-print / Author's Accepted Manuscript
Licenza: Creative commons
Dimensione 5.9 MB
Formato Adobe PDF
5.9 MB Adobe PDF Visualizza/Apri
Pubblicazioni consigliate

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11583/2989246