Mohammed Khatiri - Task scheduling on heterogeneous multi-core

Today, high-performance computing platforms (HPC) are experiencing rapid and significant development, they are bigger, faster, more powerful, but also more complex. These platforms are more and more heterogeneous, dynamic, and distributed. These characteristics create new challenges for the scheduling problem which corresponds to the allocation of tasks to the different and remote processors.
The first challenge is how to effectively manage the heterogeneity of resources which can appear at the computation level or at the communication level. The second challenge is the dynamic nature of tasks and data, To face this challenge, the development must be supported by effective software tools to manage the complexity. In this dissertation, we are interested in both on-line and off-line scheduling problems in heterogeneous resources on a dynamic environment. The crucial performance feature is communication, which is ignored in most related approaches.

Firstly, we analyze the Work Stealing on-line algorithm on parallel and distributed platforms with different contexts of heterogeneity. We start with a mathematical analysis of a new model of Work Stealing algorithm in a distributed memory platform where communications between processors are modeled by a large latency. Then, we extend the previous problem to two separate clusters, where the communication between two processors inside the same cluster is much less than external communication. We study this problem using simulations. Thus, we develop a PYTHON simulator, the simulator is used to simulate different Work Stealing algorithms in different contexts (different topologies, different tasks type, and different configurations).

In the second part of this work, we focus on two offline scheduling problems. Firstly, we consider the scheduling problem of a set of periodic implicit-deadline and synchronous tasks, on a real-time multiprocessor composed of m identical processors including communication. We propose a new task allocation algorithm that aims to reduce the number of task migrations and limits migration (of migrant tasks) on two processors. Secondly, we model a recent scheduling problem, which concerns the \textbf {micro-services} architectures which aim to divide large applications (Monolithic applications) into several micro connected applications (microservices), which makes the scheduling problem of micro-services special. Our model allows us to access several research directions able to identify effective solutions with mathematical approximations.