Resource Management Policy Handling Multiple Use-Cases in MPSoC Platforms Using Constraint Programming

摘要

Multi-processor system-on-chip (MPSoC) technology is finding widespread application in the embedded system domain, like in cell phones, automotive control units or avionics. Once deployed in field, these devices always run the same set of applications, in a well-characterized context. It is therefore possible to spend a large amount of time for off-line software optimization and then deploy the results on the field. Each possible set of applications that can be active simultaneously in an MPSoC platform leads to a different use-case that the system has to be verified and tested for. Above all, smooth switching between use-cases falls within the scope of the resource manager, since users should not experience artifacts or delays when a transition between any two consecutive use-cases takes place. In this paper, we propose a semi-static approach to the resource management problem, where the allocation and scheduling solutions for the tasks in each use-case are computed off-line via a Logic Based Benders Decomposition approach using Constraint Programming and stored for use in run-time mapping decisions. The solutions are logically organized in a lattice, so that the transition costs between any two consecutive use-cases can be bound. The resulting framework exhibits both a high level of flexibility and orders of magnitude speed ups w.r.t. monolithic approaches that do not exploit decomposition.