Optimizing edge cloud deployments for video analytics
(English) As our digital world and physical realities blend together, we, as users, are growing to expect real-time interaction wherever and whenever we want. Newer internet services require lower latency than a data center hundreds of kilometers away can provide, while generating more data than the...
| Autor: | |
|---|---|
| Formato: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2022 |
| País: | España |
| Recursos: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/692065 |
| Acesso em linha: | http://hdl.handle.net/10803/692065 https://dx.doi.org/10.5821/dissertation-2117-413896 |
| Access Level: | acceso abierto |
| Palavra-chave: | Àrees temàtiques de la UPC::Informàtica 004 |
| Resumo: | (English) As our digital world and physical realities blend together, we, as users, are growing to expect real-time interaction wherever and whenever we want. Newer internet services require lower latency than a data center hundreds of kilometers away can provide, while generating more data than the backhaul of the network can absorb. Consequently, resources are being moved to the edge of the network to create a new type of highly distributed cloud infrastructure: the edge cloud. Edge cloud deployments differ from traditional cloud deployments in two crucial points that, altogether, arise new challenges. First, edge locations are constrained by the amount of resources they can host, limiting the services these locations can execute and provide. Moreover, this limitation causes service providers to have a narrower spectrum of hardware from which to choose when executing their services at a given location while increasing the heterogeneity of the infrastructure as a whole. That is, resource-constrained nodes and hardware accelerated server-grade nodes will coexist, not within the same location, but within the same network. Services must consider this heterogeneity and adapt to it transparently. Second, contrary to traditional cloud locations, user aggregation is limited in the edge, as each location serves only its designated geographical region. Consequently, service providers face the challenge of understanding the impact of considering locality, hardware heterogeneity, and the compute requirements of their workloads while designing new deployments. Among all the services and use cases expected to be accelerated from the edge, one stands out from the rest. Video analytics is already today the main use case being deployed to the edge due to its strict latency requirements and the high amount of bandwidth it continuously generates. In fact, the edge cloud is deemed as a necessary accelerator for video analytics deployments to be feasible and cost-effective at scale. Unfortunately, video analytics is a computationally expensive task, often requiring high-end hardware acceleration to provide real-time performance to a single user. In a context in which resources become scarcer the closer they are to the edge of the network, such high computational cost easily exceeds what the infrastructure can provide unless the entire workflow is optimized. Therefore, there is a need for new techniques that can enable video analytics to be served from resource-constrained nodes without compromising the user experience, while, at the same time, being able to maximize resource utilization there where resources are scarcer. In this context, this thesis contributes to the optimization of edge cloud deployments aimed at providing service of video analytics workloads through distributed and resource-constrained edge infrastructures. Towards this end, this thesis contributes to the state-of-the-art by (C1) characterizing video analytics workloads on an heterogeneous set of hardware platform, each mapping to a different edge location archetype; (C2) developing and presenting a novel framework to accelerate large-scale deployments by leveraging the potential of a hybrid Edge-Cloud interplay and automating the task of optimizing neural networks and specializing them to the specific deployment's context; and, finally, (C3) developing a novel method to enable video analytics workloads to be massively distributed across different edge of locations. Together, the contributions of this thesis define the hardware requirements of a heterogeneous edge cloud (C1) and open the door to new ways to adapt, optimize (C2), and distribute (C3) video analytics workloads for the edge cloud. |
|---|