Section: New Results

Uniprocessor Real-Time Scheduling

Participants : Falou Ndoye, Yves Sorel, Walid Talaboulma.

Previous years, we worked on schedulability analyses of dependent tasks, executed on a uniprocessor, which take into account the exact preemption cost and more generally the cost of the real-time operating system. Indeed, this cost is composed of the cost of the preemptions and the cost of the scheduler. Our approach is based on an offline real-time schedulability analysis, proved sustainable, that produces a scheduling table. This latter contains the next instants (activation and completion of tasks) when the scheduler will be called, beeing aware of the instants where tasks are preempted and then resumed. This approach allows the schedulability analysis to account preemption costs involved by other preemptions. The scheduling table contains also the address of the next task to execute preventing the scheduler to choose it in the ready tasks list, unlike with classical on-line scheduler. The theoretical results in the uniprocessor case, are given in the Falou Ndoye's PhD thesis [19] defended in April this year. This approach has been implemented through an offline scheduler that is triggered by a timer when this latter is equal to zero and loaded with the next instant contained in the scheduling table, according to a time trigger approach.

We carried out two kinds of implementations. Actually, the first one is a simulation since our time trigger offline scheduler is modelled as a high priority task running uppon an existing operating system. We experimented this approach with Vanilla Linux (not modified) and Linux/Xenomai, real-time versions of the Linux operating system, with low latency caracteristics. Of course, these implementations were only able to show that the theoterical results were correct, but did not provide good real-time performances nor a robust way to measure time without influencing the usefull code. Therefore, we implemented our scheduler on a bare metal ARM968E-S processor based on an ARM9 architecture since it is widespread in the industry world, and we experimented this processor few years ago to determine the cost of classical online schedulers.

For this purpose we used a MCB2929 developpement board, from Keil, containing the LPC2929 SoC including the ARM968E-S, an accurate timer, and various peripherals. The scheduling table is generated offline for a set of tasks, and stored in the memory as an array of couples, each composed of the task to execute and the duration elapsing until the next scheduler call. This duration is used to set the timer counter. When it hits zero it triggers a high priority interruption that is serviced by calling again the scheduler to choose the next couple (task, duration), and so on up to the end of the scheduling table. This will repeat infinitely from the beginning of the scheduling table.

We tested different set of tasks with multiple preemption scenarios, that can yield to deadlines misses. We measured for a 12Mhz CPU and timer clock frequencies a value of 28 $\mu$s for the scheduler cost, and of 1 $\mu$s for the preemption cost.