Scheduling techniques to improve the worst-case execution time of real-time parallel applications on heterogeneous platforms

Doctoral thesis, 2021

The key to providing high performance and energy-efficient execution for hard real-time applications is the time predictable and efficient usage of heterogeneous multiprocessors. However, schedulability analysis of parallel applications executed on unrelated heterogeneous multiprocessors is challenging and has not been investigated adequately by earlier works.

The unrelated model is suitable to represent many of the multiprocessor platforms available today because a task (i.e., sequential code) may exhibit a different work-case-execution-time (WCET) on each type of processor on an unrelated heterogeneous multiprocessors platform. A parallel application can be realistically modeled as a directed acyclic graph (DAG), where the nodes are sequential tasks and the edges are dependencies among the tasks. This thesis considers a sporadic DAG model which is used broadly to analyze and verify the real-time requirements of parallel applications. A global work-conserving scheduler can efficiently utilize an unrelated platform by executing the tasks of a DAG on different processor types. However, it is challenging to compute an upper bound on the worst-case schedule length of the DAG, called makespan, which is used to verify whether the deadline of a DAG is met or not. There are two main challenges. First, because of the heterogeneity of the processors, the WCET for each task of the DAG depends on which processor the task is executing on during actual runtime. Second, timing anomalies are the main obstacle to compute the makespan even for the simpler case when all the processors are of the same type, i.e., homogeneous multiprocessors. To that end, this thesis addresses the following problem: How we can schedule multiple sporadic DAGs on unrelated multiprocessors such that all the DAGs meet their deadlines.

Initially, the thesis focuses on homogeneous multiprocessors that is a special case of unrelated multiprocessors to understand and tackle the main challenge of timing anomalies. A novel timing-anomaly-free scheduler is proposed which can be used to compute the makespan of a DAG just by simulating the execution of the tasks based on this proposed scheduler. A set of representative task-based parallel OpenMP applications from the BOTS benchmark suite are modeled as DAGs to investigate the timing behavior of real-world applications. A simulation framework is developed to evaluate the proposed method. Furthermore, the thesis targets unrelated multiprocessors and proposes a global scheduler to execute the tasks of a single DAG to an unrelated multiprocessors platform. Based on the proposed scheduler, methods to compute the makespan of a single DAG are introduced. A set of representative parallel applications from the BOTS benchmark suite are modeled as DAGs that execute on unrelated multiprocessors. Furthermore, synthetic DAGs are generated to examine additional structures of parallel applications and various platform capabilities. A simulation framework that simulates the execution of the tasks of a DAG on an unrelated multiprocessor platform is introduced to assess the effectiveness of the proposed makespan computations. Finally, based on the makespan computation of a single DAG this thesis presents the design and schedulability analysis of global and federated scheduling of sporadic DAGs that execute on unrelated multiprocessors.