Abstract
The CPU schedulers in general-purpose operating systems are designed to provide fast response time for interactive applications and high throughput for batch applications. The heuristics used to achieve these goals do not lend themselves to scheduling real-time applications, nor do they meet other scheduling requirements such as coordinating scheduling across several processors or machines, or enforcing isolation between applications, users, and administrative domains. Extending the scheduling subsystems of general-purpose operating systems in an ad hoc manner is time consuming and requires considerable expertise as well as source code to the operating system. Furthermore, once extended, the new scheduler may be as inflexible as the original.
The thesis of this dissertation is that extending a general-purpose operating system with a general, heterogeneous scheduling hierarchy is feasible and useful. A hierarchy of schedulers generalizes the role of CPU schedulers by allowing them to schedule other schedulers in addition to scheduling threads. A general, heterogeneous scheduling hierarchy is one that allows arbitrary (or nearly arbitrary) scheduling algorithms throughout the hierarchy. In contrast, most of the previous work on hierarchical scheduling has imposed restrictions on the schedulers used in part or all of the hierarchy.
This dissertation describes the Hierarchical Loadable Scheduler (HLS) architecture, which permits schedulers to be dynamically composed in the kernel of a general-purpose operating system. The most important characteristics of HLS, and the ones that distinguish it from previous work, are that it has demonstrated that a hierarchy of nearly arbitrary schedulers can be efficiently implemented in a general-purpose operating system, and that the behavior of a hierarchy of soft real-time schedulers can be reasoned about in order to provide guaranteed scheduling behavior to application threads. The flexibility afforded by HLS permits scheduling behavior to be tailored to meet complex requirements without encumbering users who have modest requirements with the performance and administrative costs of a complex scheduler.
Contributions of this dissertation include the following. (1) The design, prototype implementation, and performance evaluation of HLS in Windows 2000. (2) A system of guarantees for scheduler composition that permits reasoning about the scheduling behavior of a hierarchy of soft real-time schedulers. Guarantees assure users that application requirements can be met throughout the lifetime of the application, and also provide application developers with a model of CPU allocation to which they can program. (3) The design, implementation, and evaluation of two augmented CPU reservation schedulers, which provide increase scheduling predictability when low-level operating system activity steals time from applications.
