Symphony: A Scalable Emulator for Distributed Systems

• Autor(a) principal: Atalaia, André de Oliveira
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/160382
• topic: Large Scale Distributed systems evaluation; network emulation; container technology; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Distributed Systems are becoming increasingly prevalent in our daily lives and their rele- vance has never been so prominent in everyday activities. This naturally leads systems to increase in complexity and dimension, with users having much higher expectations about the performance and correctness of these systems. This creates a significant pressure for conducting systematic and detailed performance assessments of novel proposals through experimentation. Conducting experiments on distributed systems can be challenging, since one has to launch multiple processes and retain control over them. While you can manually run and control experiments for small distributed systems, such possibility becomes daunting when large-scale distributed systems such as peer-to-peer systems or recent proposes in the context of Web 3.0. To be able to extract key performance metrics, which depend heavily on the underlying network, and since the networks where these systems operate are many times unstable, this can lead to inconclusive results. Moreover, when considering systems operating at a global scale, it is extremely hard to have access to competitive resources in many locations, to achieve realistic conditions. To emulate the underlying network realistically we can use Linux tc to shape links between the end hosts to exhibit target properties. To do this for large systems with thousands of nodes is however, very complex. In this work, we introduce a tool that assists in defining and executing experiments for large scale distributed systems, using container technology and network emulation. We present an architecture where a manager node exists per physical machine, and this man- ager is responsible for all the application nodes in that physical machine. This responsi- bility goes from receiving the experiment information (network topology, link properties, dynamic events), to ensuring that every application node receive their corresponding events. Lastly, we evaluated our tool by conducting experiments and comparing it against state-of-the-art tools, evaluating their performance when emulating large distributed systems.