Extending a Best-Effort Operating System to Provide QoS Processor Management

The benefits of QoS network features are easily lost when the endnodes are managed by a conventional, best-effort operating system. Schedulers of such operating systems provide only rudimentary tools (like priority adjustment) for processor management. We present here a simple extension to a processor management system that allows an application to reserve a share of the processor for a specified interval. The system is targeted at applications with frequently changing resource demands or recurring, though non-periodic resource requests. An example of such an application is a network-aware image search and retrieval system, but other network-aware client-server applications also fall into the same category. The admission control component of the processor management system decides if a resource request can be satisfied. To limit the amount of time spent negotiating with the operating system, the application can present a ranked list of acceptable reservations. The admission controller then picks the best request that can still be satisfied (using the Simplex linear programming algorithm to find the best solution). If there are insufficient resources, the application must deal with the shortage. Any possible adaptation (if the accepted request was not the application’s first choice) is left to the application. The processor management system has been implemented for NetBSD and been ported to Linux, and the paper includes an evaluation of its effectiveness. The overhead is low, and although reservations are not guaranteed, in practical settings the application almost always obtains the cycles requested.