Martijn van den Heuvel
Many industrial systems have various hardware and software functions for controlling mechanics. If these functions act independently, as they do in legacy situations, their overall performance is not optimal. There is a trend towards optimizing the overall system performance and creating a synergy between the different functions in a system, which is achieved by replacing more and more dedicated, single-function hardware by software components running on programmable platforms. This increases the re-usability of the functions, but their synergy requires also that (parts of) the multiple software functions share the same embedded platform.
In this work, we look at the composition of inter-dependent software functions on a shared platform from a timing perspective. We consider platforms comprised of one preemptive processor resource and, optionally, multiple non-preemptive resources. Each function is implemented by a set of tasks; the group of tasks of a function that executes on the same processor, along with its scheduler, is called a component. The tasks of a component typically have hard timing constraints. Fulfilling these timing constraints of a component requires analysis. Looking at a single function, co-operative scheduling of the tasks within a component has already proven to be a powerful tool to make the implementation of a function more predictable. For example, co-operative scheduling can accelerate the execution of a task (making it easier to satisfy timing constraints), it can reduce the cost of arbitrary preemptions (leading to more realistic execution-time estimates) and it can guarantee access to other resources without the need for arbitration by other protocols. Since timeliness is an important functional requirement, (re-)use of a component for composition and integration on a platform must deal with timing.
To enable us to analyze and specify the timing requirements of a particular component in isolation from other components, we reserve and enforce the availability of all its specified resources during run-time. The real-time systems community has proposed hierarchical scheduling frameworks (HSFs) to implement this isolation between components. After admitting a component on a shared platform, a component in an HSF keeps meeting its timing constraints as long as it behaves as specified. If it violates its specification, it may be penalized, but other components are temporally isolated from the malignant effects. A component in an HSF is said to execute on a virtual platform with a dedicated processor at a speed proportional to its reserved processor supply. Three effects disturb this point of view. Firstly, processor time is supplied discontinuously. Secondly, the actual processor is faster. Thirdly, the HSF no longer guarantees the isolation of an individual component when two arbitrary components violate their specification during access to non-preemptive resources, even when access is arbitrated via well-defined real-time protocols.
The scientific contributions of this work focus on these three issues. Our solutions to these issues cover the system design from component requirements to run-time allocation. Firstly, we present a novel scheduling method that enables us to integrate the component into an HSF. It guarantees that each integrated component executes its tasks exactly in the same order regardless of a continuous or a discontinuous supply of processor time. Using our method, the component executes on a virtual platform and it only experiences that the processor speed is different from the actual processor speed. As a result, we can focus on the traditional scheduling problem of meeting deadline constraints of tasks on a uni-processor platform. For such platforms, we show how scheduling tasks co-operatively within a component helps to meet the deadlines of this component. We compare the strength of these cooperative scheduling techniques to theoretically optimal schedulers.
Secondly, we standardize the way of computing the resource requirements of a component, even in the presence of non-preemptive resources. We can therefore apply the same timing analysis to the components in an HSF as to the tasks inside, regardless of their scheduling or their protocol being used for non-preemptive resources. This increases the re-usability of the timing analysis of components. We also make non-preemptive resources transparent during the development cycle of a component, i.e., the developer of a component can be unaware of the actual protocol being used in an HSF. Components can therefore be unaware that access to non-preemptive resources requires arbitration. Finally, we complement the existing real-time protocols for arbitrating access to non-preemptive resources with mechanisms to confine temporal faults to those components in the HSF that share the same non-preemptive resources. We compare the overheads of sharing non-preemptive resources between components with and without mechanisms for confinement of temporal faults. We do this by means of experiments within an HSF-enabled real-time operating system.