Date of Award
Doctor of Philosophy (PhD)
In order to provide safety guarantees or quality of service guarantees, many of today's systems consist of latency-critical applications, e.g. applications with timing constraints. The problem of scheduling multiple latency-critical jobs on a multiprocessor or multicore machine has been extensively studied for sequential (non-parallizable) jobs and different system models and different objectives have been considered. However, the computational requirement of a single job is still limited by the capacity of a single core. To provide increasingly complex functionalities of applications and to complete their higher computational demands within the same or even more stringent timing constraints, we must exploit the internal parallelism of jobs, where individual jobs are parallel programs and can potentially utilize more than one core in parallel. However, there is little work considering scheduling multiple parallel jobs that are latency-critical.
This dissertation focuses on developing new scheduling strategies, analysis tools, and practical platform design techniques to enable efficient and scalable parallel real-time scheduling for latency-critical applications on multicore systems. In particular, the research is focused on two types of systems: (1) static real-time systems for tasks with deadlines where the temporal properties of the tasks that need to execute is known a priori and the goal is to guarantee the temporal correctness of the tasks prior to their executions; and (2) online systems for latency-critical jobs where multiple jobs arrive over time and the goal to optimize for a performance objective of jobs during the execution.
For static real-time systems for parallel tasks, several scheduling strategies, including global earliest deadline first, global rate monotonic and a novel federated scheduling, are proposed, analyzed and implemented. These scheduling strategies have the best known theoretical performance for parallel real-time tasks under any global strategy, any fixed priority scheduling and any scheduling strategy, respectively. In addition, federated scheduling is generalized to systems with multiple criticality levels and systems with stochastic tasks. Both numerical and empirical experiments show that federated scheduling and its variations have good schedulability performance and are efficient in practice.
For online systems with multiple latency-critical jobs, different online scheduling strategies are proposed and analyzed for different objectives, including maximizing the number of jobs meeting a target latency, maximizing the profit of jobs, minimizing the maximum latency and minimizing the average latency. For example, a simple First-In-First-Out scheduler is proven to be scalable for minimizing the maximum latency. Based on this theoretical intuition, a more practical work-stealing scheduler is developed, analyzed and implemented. Empirical evaluations indicate that, on both real world and synthetic workloads, this work-stealing implementation performs almost as well as an optimal scheduler.
Chenyang Lu, Kunal Agrawal
Sameh Elnikety, Christopher Gill, Roch Guerin, I-Ting Angelina Lee