Section: New Results

Multiprocessor Real-Time Scheduling

Participants : Abderraouf Benyahia, Laurent George, Mohamed Marouf, Falou Ndoye, Simon Nivault, Yves Sorel, Cécile Stentzel, Meriem Zidouni.

Non-Preemptive Partitioned Fault Tolerant Scheduling

We addressed partitioned multiprocessor scheduling of non-preemptive strict periodic tasks which is extended thereafter to hardware fault tolerance [17] .

In order to schedule a task set of non-preemptive strict periodic tasks on a multiprocessor platform, we partitioned this task set into subsets of tasks, each one is scheduled on a single processor using our proposed uniprocessor scheduling algorithm. The partition is carried out according to an enhanced “First Fit” algorithm that balances the load of the tasks on all the processors. However, inter-processors communications can lead to delay task execution. Thus, we determined the start time of each task taking into account the communication delay between this latter task and its predecessor tasks. Also, as inter-processor communications may generate a transient phase, we computed the length of the transient phase.

We proposed a fault tolerant real-time scheduling algorithm which allows hardware processors and/or buses faults, and conserves the strict periodicity of each task. We also proposed a graph transformation algorithm, applied on the task graph, which generates redundancies of tasks as well as dependencies. The transformation adds also selector tasks which choose data coming from the non failing processors and buses. That algorithm is based on exclusion relations to assign redundant tasks (resp. dependencies) to different processors (resp. busses). Then, we extended the previous partitioned multiprocessor scheduling algorithm to manage fault tolerance taking into account these exclusion relations.

This approach was successfully implemented on a CyCabs electric vehicle in a real-time fault tolerant tracking application where some processor or some bus could fail without any consequence on the proper execution of the application, i.e. same functional behaviour and real-time constraints satisfied.

Partitioned Scheduling with Exact Preemption Cost

Preemption allows a better scheduling success ratio but has a cost that must not be neglected in safety critical applications of domains such as avionic, automotive, etc. We focused on partitioned multiprocessor scheduling of independent preemptive periodic real-time tasks, while taking into account the exact preemption cost with the $\oplus$ operation formerly proposed by Meumeu and Sorel [10] . We improved the “greedy” heuristic proposed last year and compared it with the “Best-Fit” (BF) and “Worst-Fit” (WF) heuristics classically used in partitioned multiprocessor scheduling, but extended to take into account the exact preemption cost. We also compared our heuristic with an exact “Branch and Bound” algorithm with the same extension. The first comparison shows that the task allocation found by our heuristic gives a better response time than those found by WF and BF. This is due to the fact that the execution of the tasks is better parallelized. On the other hand, BF and WF heuristics execute a bit faster than our heuristic because they do not use all the available processors contrary to our heuristic which has the advantage to improve the load balancing of the tasks on all the processors.

Then, we addressed the scheduling of preemptive periodic real-time tasks with dependence constraints involving task precedences and data dependences. We considered harmonic tasks, i.e. periods of tasks are multiple or equal, to avoid loss of data. In order to satisfy data dependence constraints, we modified the release dates and deadlines of the dependent tasks according to the reception date of the data. In addition, data dependences between tasks mean to share data between dependent tasks which can cause deadlock and priority inversion problems. In order to solve these problems while taking into account the preemption cost, we proposed a new schedulability condition based on an extension of the $\oplus$ operation. We plan to propose a mutiprocessor scheduling heuristic based on that condition applied on tasks with modified release dates and deadlines.

Semi-partitioned Scheduling

Semi-partitioned multiprocessor scheduling stands between partitioned and global scheduling, the latter allowing migrations. We mainly addressed the semi-partitioned scheduling approach where the Worst Case Execution Time (WCET) of a job can be portioned, each portion being executed on a dedicated processor, according to a static pattern of migration. A job is migrated at its local deadline, computed from the deadline of the task it belongs to. We have studied this approach in the context of a fork/join task model with thread parallelism. A task is composed of a sequence of segments that can be parallelized in threads, if needed. The local deadlines depends on the number of parallel threads assigned to each segment.

Code Generation for Multicore

This work was carried out in the OPENPROD ITEA project 8.3.2.2 . xMod developed by IFPEN (IFP Energies Nouvelles), is an heterogeneous model integration environment that allows model importation from specific tools such as Simulink, AMSIM, etc. It also provides as a virtual instrumentation laboratory. In order to make xMod being able to run simulations with hardware-in-the-loop environment, we developed a new SynDEx executive kernel based on the kernel, dedicated to Windows/RTX, developed last year. That executive kernel is used with the macro-code generated by SynDEx to produce a real-time executable code that can drive the execution (real-time multi-core distribution and synchronized execution) of the models imported by xMod and simulated in the virtual instrumentation laboratory. This prototype as well as the report describing the corresponding achieved works, are the final deliverable of the OPENPROD project.

Furthermore, a French and English SynDEx code generation reference manual has been written to help future SynDEx users and maintainers to generate real-time code for already supported architectures or new ones.

Gateway with Modeling Languages for Certified Code Generation

This work was carried out in the P FUI project 8.2.2.1 . We provide inside the project expertise mainly on schedulability analysis and automatic generation of distributed real-time code. In this context, we developed a gateway between UML/MARTE and SynDEx. From a model specified with UML (Activity Diagram to specify algorithms and Composite Structure Diagram to specify multicomponent architectures) and refined with the UML profile MARTE (Modeling and Analysis of Real-Time Embedded Systems), we use the gateway to generate automatically distributed real-time application specified in the SynDEx format. Currently, we intend to provide a gateway between the GeneAuto language and SynDEx. The GeneAuto language is a subset of the future pivot P language. We presently deal with the part of the GeneAuto language corresponding to Simulink for data-flow modeling and we plan to deal soon with the part corresponding to Stateflow for control-flow modeling (composition of automata).

SynDEx Updates

We continued the software developments for the future version 8 of SynDEx which will feature a new software architecture to allow better functionality evolutions and maintenance. On the other hand in the COTROS ADT ("Génération de code temps réel distribué optimisé et sûr"), we completed the tests on the new automatic code generator for the current version 7 of SynDEx. This new generator produces code for mono-periodic and multi-periodic applications with condition and repetitive control structures, for the different hardware architectures supported by SynDEx. We developed a checker for the generated code that was integrated in the new generator. This checker verifies the correct use of semaphores and consequently the absence of deadlocks in the real-time code. Deadlocks are the most difficult part when dealing with distributed architectures. We achieved also a maintenance report describing the structure and the main features of code generator, as well as the technical choices we did.