Efficient Execution of Dependent Tasks on Many-Core Processors

International audienceThe increasing performance requirements of safety-critical real-time embedded systems made traditional single-corearchitectures obsolete. Moving to more complex many-core systems requires new techniques and tools for the certificationof the embedded software. Timing and functional behaviours are subject to specific requirements of certification guidelinessuch as D-178B/C for avionics and ISO26262 for automotive systems. Determinism and predictability of such systems are amajor challenge. Running tasks should be known in advance which makes the static and non-preemptive scheduling a suitableapproach to reach an optimal execution with a guarantee of a certain degree of determinism. Recent work on mapping andscheduling problems [1], [2], [3] consider a known value (or a set of possible values) of the Worst-Case Response Time (WCRT)and computes a mapping that optimizes a predefined cost function. When the response time analysis is too pessimistic, thestatic scheduling may introduce an idle time which reduces the core utilization. Scheduling techniques must rely on tight estimations of the WCRT which in turn depends on co-runner tasks. However, inorder to obtain a tight upper-bound on the response time, a mapping and scheduling should be known in advance. Indeed, theresponse time is highly influenced by the co-runner tasks. Concurrent accesses to the same shared resource may introduceinterferences that should be accounted for in the response time analysis. The search for an optimal scheduling with a tightWCRT analysis that includes the shared resource interferences is a challenging open problem