PhD Thesis

< back to PhD thesis

« Study of task migration in a multi-tiled architecture - Automatic generation of an agent-based solution ».

Author: A. El-Antably
Advisor: F. Rousseau
President of jury: J.-F. Méhaut
thesis reviewer(s): G. Sassatelli, F. Hessel,
thesis examinator(s): O. Gruber, F. De Dinechin,
These de Doctorat Université de Grenoble
Speciality: Nanoélectronique et Nanotechnologies
Defense: December 16 2015
ISBN: 978-2-11-129207-9


Fully distributed memory multi-processors (MPSoC) implemented in multi-tiled architectures are promising solutions to support modern sophisticated applications, however, reliability of such systems is always an issue. As a result, a system-level solution like task migration keeps its importance. Transferring the execution of a task from one tile to another helps keep acceptable reliability of such systems. A tile contains at least one processor, private main memory and associated peripherals with a communication device responsible for inter-tile communications. We propose in this work an agent based task migration technique that targets data-flow applications running on multi-tiled architectures. This technique uses a middleware layer that makes it transparent to application programmers and eases its portability over different multi-tiled architectures. In order for this solution to be scalable to systems with more tiles, an automatic generation tool-chain is designed to generate migration agents and provide them with necessary information enabling them to execute migration processes properly. Such information is extracted automatically from application(s) task graphs and mapping on the system tiles. We show how agents are placed with applications and how such necessary information is generated and linked with them. The tool-chain is capable of generating code for ARM and x86 architectures. This task migration technique can be deployed on small operating systems that support neither MMU nor dynamic loading for task code. We show that this technique is operational on x86 based real hardware platform as well as on an ARM based simulation platform. Experimental results show low overhead both in memory and performance. Performance overhead due to migration of a task in a typical small application where it has one predecessor and one successor is 18.25%.

pdf pdf