Uma proposta dirigida por dados para a melhoria do engajamento e da alocação de recursos em transmissões adaptativas ao vivo

Detalhes bibliográficos
Ano de defesa: 2023
Autor(a) principal: Thiago Amaral Guarnieri
Orientador(a): Não Informado pela instituição
Banca de defesa: Não Informado pela instituição
Tipo de documento: Tese
Tipo de acesso: Acesso aberto
Idioma: por
Instituição de defesa: Universidade Federal de Minas Gerais
Brasil
ICX - DEPARTAMENTO DE CIÊNCIA DA COMPUTAÇÃO
Programa de Pós-Graduação em Ciência da Computação
UFMG
Programa de Pós-Graduação: Não Informado pela instituição
Departamento: Não Informado pela instituição
País: Não Informado pela instituição
Palavras-chave em Português:
QoS
QoE
Link de acesso: http://hdl.handle.net/1843/58284
Resumo: Internet live streaming has reached large audiences. However, with the rise in popularity, fewer infrastructure resources become available to meet each user performance requirements. In other words, it becomes harder to conciliate high transmission performance and transmission scale growth. One of the approaches to reacting to resource constraints and maintaining a minimum client transmission performance is the use of adaptation mechanisms, which adjusts the bitrate to the client device type and bandwidth. This dynamic bitrate adaptation reduces the probability of reproduction stalls, which have a negative perception by the users. On the other hand, the content provider needs to keep the system available for new users and uses for this goal resource allocation plans, which reduces, when necessary, the bitrate of its clients. The bitrate reduction allows the entrance of new clients. However, it can produce a negative impact on the current users. As a result, they end by abandoning their sessions. In other words, the video bitrate reduction leads to a user engagement reduction. Therefore, there is a conflict of interest where the user always wants the maximum possible bitrate, and the content provider wants to maximize user engagement in both the number of users and client session duration, which may require client bitrate reduction. Based on this perspective, this thesis has as main objectives: (1) to contribute to the current literature concerning the relationship between client transmission performance and engagement. This knowledge allows the creation of engagement and client behavior models that help content providers in infrastructure planning and performance monitoring, and (2) to explore, through these models, resource allocation alternatives to achieve a better tradeoff between user and content provider interests, that is, to increase resource saving in provider while it preserves engagement of the current users. The path to reaching this better tradeoff is the creation of personalized resource limitations that considers each client’s transmission performance requirements. This thesis map these expressed objectives in four research questions as follows: (1) to characterize client transmission performance in large-scale live streaming and the correlation of this performance with user engagement; (2) to develop a client behavior model that considers the impact of the client transmission performance on user engagement and the client adaptation regime; (3) to develop engagement descriptive and predictive models for active monitoring of the engagement in live video streaming a and (4) to project a mechanism for content provider resource allocation that evaluates various allocation sce- narios to choose the most suitable for each client individually to preserve user engagement and reduce content provider resource consumption. The main contributions of this thesis are: (1) a characterization of client transmission performance in a large-scale event with millions of simultaneous users and and evaluation of the impact of this performance on user engagement. We propose a concept of performance scenarios that show that the tolerance to a variation in a particular performance metric varies depending on the value of other performance metrics. For example, the rise in the client bitrate increases user engagement only if the stall and adaptation rates are low. We also addressed the impact of contextual factors on performance metrics and engagement. We found that the device type, platform, internet service provider, and transmission period influence client transmission performance and engagement. Besides this, we also investigated the impact of transmission scale on transmission performance. We verify that the transmission infrastructure applied bitrate limitations to deal with heavy workloads; (2) the creation and validation of a performance-aware client behavior model. This model revealed that client transmission performance impacts user engagement (permanence, time between sessions, and the number of sessions) and client adaptation regime; (3) the creation of descriptive and predictive engagement models that advance the state-of-the-art concerning precision and accuracy. These models introduce a new approach to describe client performance. Instead of using classical performance metrics like stall and adaptation rate, we used the client adaptation regime, stored in a transition matrix, associated with performance scenarios. Using this strategy, we constructed specialized models capable of reaching high accuracy in different client transmission performance levels. More specifically, the descriptive model has reached accuracy nearly 90% against 65% of the engagement model that uses classical performance metrics. The predictive model, in turn, obtained an 80% acuraccy in association with performance scenarios; and (4) The proposition of a resource allocation mechanism that considers the impact of the adaptation decisions on user engagement. We use the preservation of user engagement to guide allocation decisions which ensures, at the same time, user performance requirements and the reduction of client resource consumption. Using trace-based simulation, we verified an average gain of 100% in user engagement and a rise of hundreds of clients every minute. Considering the resource-saving mode, we registered a reduction of hundreds of gigabytes in bandwidth usage, with an impact of 0.4% in the original engagement.