Processor Capacity Reserves: An Abstraction for Managing Processor Usage

Multimedia applications require operating systems that support time-constrained data types such as digital audio and video. These continuous media [1] demand timely service from the system, and time-sharing scheduling algorithms are not sufficient. Furthermore, simple fixed priority scheduling, used in many hard real-time systems, does not necessarily guarantee the successful execution of arbitrary collections of programs which may have conflicting timing requirements or which may overload the system [7]. We have designed and implemented a processor capacity reservation mechanism to allow programs to reserve the capacity they need to run. Enforcement of reservations provides a scheduling firewall to protect programs from competition for the processor; this is similar to the firewall provided by memory protection, which isolates programs from outside interference in their address spaces. Using reservation, a system can control processor capacity allocation in the same way that it controls discrete resources like memory or storage space, and this prevents over-committing the processor. Our reservation mechanism depends on a scheduling framework where each reservation is expressed as a rate of progress (defined as computation time per period of real time). A rate can be associated with non-periodic programs as well as periodic programs. An admission control policy, based on rate monotonic scheduling theory [5, 6], moderates access to processor resources. An enforcement mechanism ensures that the reserved computation is limited to the requested rate and does not interfere with other activities. The reservation mechanism measures the processor usage of each reserved program and even includes in that measurement any computation time consumed by servers invoked on that program’s behalf. A detailed description of the scheduling framework, the admission control policy, and the enforcement mechanism appears elsewhere [7]. Supporting multimedia applications is our primary goal in this work. Many multimedia systems assume that this kind of processor reservation functionality is available [1, 2], but previous reservation systems typically use extended time-sharing algorithms which attempt to match long-term utilization with preset reservation values [3, 4]. These systems do not have the accurate usage measurement and control required by low-latency multimedia applications. A reservation mechanism like ours is also useful in other application areas. Multiprocessor systems can use reservations to interleave parallel programs on a collection of processors. The accurate processor