Sever Institute of Technology Department of Computer Science Flexible Scheduling in Middleware for Distributed Rate-based Real-time Applications

FLEXIBLE SCHEDULING IN MIDDLEWARE FOR DISTRIBUTED RATE-BASED REAL-TIME APPLICATIONS by Christopher D. Gill ADVISOR: Dr. Ron K. Cytron and Dr. Douglas C. Schmidt May, 2002 Saint Louis, Missouri Distributed rate-based real-time systems, such as process control and avionics mission computing systems, have traditionally been scheduled statically. Static scheduling provides assurance of schedulability prior to run-time and can be implemented with low run-time overhead. However, static scheduling is brittle in the face of unanticipated overload, and treats invocation-to-invocation variations in resource requirements inflexibly. As a consequence, processing resources are often under-utilized in the average case, and the resulting systems are hard to adapt to meet new real-time processing requirements. Dynamic scheduling offers relief from the limitations of static scheduling. However, dynamic scheduling often has a higher run-time cost because certain decisions are enforced on-line. Furthermore, under conditions of overload tasks can be scheduled dynamically that may never be dispatched, or that upon dispatch would miss their deadlines. We review the implications of these factors on rate-based distributed systems, and posits the necessity to combine static and dynamic approaches to exploit the strengths and compensate for the weaknesses of either approach in isolation. We present a general hybrid approach to real-time scheduling and dispatching in middleware, that can employ both static and dynamic components. This approach provides (1) feasibility assurance for the most critical tasks, (2) the ability to extend this assurance incrementally to operations in successively lower criticality equivalence classes, (3) the ability to trade off bounds on feasible utilization and dispatching overhead in cases where, for example, execution jitter is a factor or rates are not harmonically related, and (4) overall flexibility to make more optimal use of scarce computing resources and to enforce a wider range of application-specified execution requirements. This approach also meets additional constraints of an increasingly important class of rate-based systems, those with requirements for robust management of realtime performance in the face of rapidly and widely changing operating conditions. To support these requirements, we present a middleware framework that implements the hybrid scheduling and dispatching approach described above, and also provides support for (1) adaptive re-scheduling of operations at run-time and (2) reflective alternation among several scheduling strategies to improve real-time performance in the face of changing operating conditions. Adaptive re-scheduling must be performed whenever operating conditions exceed the ability of the scheduling and dispatching infrastructure to meet the critical real-time requirements of the system under the currently specified rates and execution times of operations. Adaptive re-scheduling relies on the ability to change the rates of execution of at least some operations, and may occur under the control of a higherlevel middleware resource manager. Different rates of execution may be specified under different operating conditions, and the number of such possible combinations may be arbitrarily large. Furthermore, adaptive re-scheduling must occur within an acceptably predictable and narrow interval, while preserving schedulability assurances for critical operations, and optimizing other properties such as resource utilization to the extent possible. To address these constraints, we describe extensions to provide flexible and efficient strategies for rate selection and priority re-assignment. Unfortunately, adaptive re-scheduling may in turn require notification of ratesensitive application components. It is therefore desirable to handle variations in operating conditions entirely within the scheduling and dispatching infrastructure when possible. A rate-based distributed real-time application, or a higher-level resource manager, could thus fall back on adaptive re-scheduling only when it cannot achieve acceptable real-time performance through self-adaptation. Reflective alternation among scheduling heuristics offers a way to tune real-time performance internally, and we offer foundational support for this approach. In particular, run-time observable information such as that provided by our metrics-feedback framework makes it possible to detect that a given current scheduling heuristic is underperforming the level of service another could provide. Furthermore we present empirical results for our framework in a realistic avionics mission computing environment. This forms the basis for guided adaptation. This dissertation makes five contributions in support of flexible and adaptive scheduling and dispatching in middleware. First, we provide a middleware scheduling framework that supports arbitrary and fine-grained composition of static/dynamic scheduling heuristics, to assure critical timeliness constraints while improving noncritical performance under a range of conditions. Second, we provide a flexible dispatching infrastructure framework composed of fine-grained primitives, and describe how appropriate configurations can be generated automatically based on the output of the scheduling framework. Third, we describe algorithms to reduce the overhead and duration of adaptive rescheduling, based on sorting for rate selection and priority assignment. Fourth, we provide timely and efficient performance information through an optimized metrics-feedback framework, to support higher-level reflection and adaptation decisions. Fifth, we present the results of empirical studies to quantify and evaluate the performance of alternative canonical scheduling heuristics, across a range of load and load jitter conditions. These studies were conducted within an avionics mission computing application framework running on realistic middleware and embedded hardware. The results obtained from these studies (1) demonstrate the potential benefits of reflective alternation among distinct scheduling heuristics at run-time, and (2) suggest performance factors of interest for future work on adaptive control policies and mechanisms using this framework.