AbstractsComputer Science

Simulating run-time task migration in many-core systems

by Koosha Ahmadi




Institution: University of New South Wales
Department: Computer Science & Engineering
Year: 2014
Keywords: Mapping; Task Migration; Many-Core Systems; Simulator
Record ID: 1051168
Full text PDF: http://handle.unsw.edu.au/1959.4/54223


Abstract

With the ever increasing number of processing elements on Network-on-Chip multiprocessors, it is crucial to utilise the resources efficiently. Allocation of tasks to processing elements is one of the most important problems to be solved for these systems. This requirement is further complicated as embedded systems become evermore open ended. In addition, as the requirements and the volume of applications are hard to predict at design time, not only do we need to allocate resources at runtime, but we may want to reallocate them as well. Having frameworks for testing different allocation and reallocation policies is therefore of great importance. However, simulators that are readily available for performing task migration and dynamic mapping in the context of open-ended embedded systems are scarce or hard to configure. We propose a framework for testing different allocation and reallocation policies for 2D-mesh many-core systems, where requirements of applications are not known a priori. Our simulator is based on the Booksim2 network simulator, and we have used Synchronous Data Flow Graphs provided by the SDF3 tool as our application model. The presented simulator is extensible and provides useful insights into on-chip communications and runtime behaviour of applications. It is therefore to be hoped that it could be used to explore and develop task migration policies.