Nicholas Mc Guire, DSLab, Lanzhou University, China

Overall complexity of computing systems is increasing with the consequences that deterministic properties of RTOS are increasingly hard (or impossible) to quarantee. The basic tenet of RTOS design in the past has been keep-it-simple but this is changing for a number of reasons - multi-core being one of them. With RT_PREEMPT maybe being one of the most complex RTOS on the planet it is a good candidate to look into radically different ways of approaching RTOS. Rather than minimizing complexity and trying to achieve macroscopic determinism by determinism at the primitive instruction level along with well defined (and known to be bounded) paths of execution - we propose to view an RTOS as a set of functional primitives that only exhibit well defined execution time distributions and treat independence with respect to the actual execution path taken, on each demand, as the critical property to ensure.

In the 1950s aircraft were built with a relatively simple system view - if each of the components was built to the highest reliability standards possible then the aggregated aircraft would be sufficiently reliable. The problem with this though is that the reliability of linear compositions decline exponentially - so this approach was just a matter of time until it would fail horribly - which it did by the mid 1950s. The mitigation was to utilize architectural means so as to achieve an increased reliability of the assembly provided the failures were independent of each other. Precisely this is the approach taken here based on the above described change of viewpoint of what an RTOS design needs to focus on: We simply propose to exchange temporal boundedness of the individual function/execution path by a well behaved distribution independence of chosen path and employment of architectural protection means.

The basic concept of replicaRT is to allow multiple identical instances of functions to concurrently execute on different cores of a multi-core processor and at points where the internal state of the function is to be externalized evaluate which of the concurrent instances "won" the race. If the execution- time distribution is well behaved and the actual execution path taken are sufficiently independent of each other, then the real-time properties of the system with respect to periodic/timed task execution jitter can be guaranteed with narrow bounds at a very high confidence level.

In this paper we outline the basic concept of replicaRT as well as a first somewhat naive model and proposal an abstract SCHED_REPLICA to be considered as an alternative real-time scheduling class for multi-core systems and present early results of "hand-optimized" replicated RT threads on NN Hardware (ARM,X86).