Advancing multi-core scheduling for real-time embedded systems

Abstract

With the widespread of multi-core architectures, real-time multi-core scheduling research has been steadily gaining importance. In spite of some remarkable achievements in real-time multi-core scheduling in the past, there are still significant research challenges that remain. In particular, the vast majority of existing research considers simple scheduling models, such as sequential tasks under RM and EDF scheduling, originally developed with uni-core platforms. Therefore, there is a need to take into account more general scheduling models that can take full advantage of multi-core processing. In this dissertation, we aim to apply successful results to more general scheduling models. To this end, we understand the fundamental differences between each scheduling model and introduce novel concepts that explain the differences. Building upon them, we develop new techniques that can easily extend existing approaches to general scheduling models, significantly advancing the state-of-the-art scheduling algorithms and schedulability analysis for real-time multi-core systems.

For a scheduler component, although job-level priority assignment (JPA) is a generalization of task-level priority assignment (TPA), there is no clear dominance relation between existing TPA and JPA scheduling techniques. We focus on understanding fundamental differences between TPA and JPA and derive a condition under which JPA is able to behave exactly the same way as TPA. Then, we present the first algorithm (SPDF- Smallest Pseudo-Deadline First) that generalizes TPA scheduling to JPA scheduling with a simple task-level parameter called pseudo-deadline. We show that SPDF dominates all TPA algorithms when pseudo-deadlines are appropriately determined.

For a task component, although a number of studies extensively investigated the schedulability analysis of sequential tasks with many influential results, the insights behind those successful results are not directly applicable or easily extensible to parallel tasks. We extend the notion of interference to capture thread-level parallelism for parallel tasks. We then leverage the proposed notion of parallelism-aware interference to derive the first schedulability test that is directly applicable to parallel tasks on multi-core platforms.

For a resource component, almost all scheduling approaches for heterogeneous multi-core platforms have been focused on partitioned scheduling rather than global scheduling, due to difficulties of applying existing work on homogeneous platforms to global heterogeneous multi-core scheduling. We extend scheduling guidelines and algorithms for optimal homogeneous multi-core scheduling toward two-type heterogeneous global scheduling to develop the first optimal scheduling algorithm for two-type heterogeneous multi-core platforms.