Otimizando a execução de aplicações paralelas em ambiente de nuvem heterogênea

Abstract

Cloud computing is already a well-established platform for high-performance application execution, among other factors, due to its resource scalability and high availability. At the same time, updating computational nodes in these systems can lead to resource heterogeneity. In this sense, the challenge of executing parallel applications in the cloud is not only related to defining the best number of threads for the application, but also to choosing the ideal architecture that will execute such application. Therefore, in this dissertation, we present the TLP-Allocator, a framework to optimize the execution of parallel applications in a heterogeneous cloud environment based on the ideal TLP degree + target architecture, implemented on top of Kubernetes technology. The TLP-Allocator is completely transparent to end-users and clients. Given a set of parallel applications to be executed, it finds the ideal TLP degree for each application and performs a match between application and ideal architecture to better utilize the available resources. We showed that analyzing the scalability impact of parallel applications on nodes with different computational processing power capabilities can bring significant gains in performance and energy consumption when combined with the TLP-Allocator. By considering the degree of parallelism of an application and the characteristics of the computational node, it is possible to obtain a more efficient and optimized allocation of the available resources, thus achieving the best EDP (Energy-Delay product) for applications with different degrees of parallelism.