Operating system/application concurrency in tightly-coupled multiple-processor systems
James W. Wendorf · 1987
Operating system processing often accounts for 30% or more of a system's total computation time. Special hardware support for OS functions can significantly improve the OS and overall system performance. A common hardware support technique is to add specialized processors to a computer system. However, previous studies have not attempted to separate the benefits of the resulting OS/application concurrency from the effects of faster, specialized hardware. To test the hypothesis that significant performance improvements can be achieved through concurrency alone, our studies focus on software-level functional specialization (SLFS), wherein general purpose processors are added to a host computer system to support OS processing concurrently with host application processing. The potential benefits of SLFS over specialized hardware include reduced development costs, increased flexibility, and increased reliability through function migration. This thesis first presents a model that defines OS/application concurrency precisely, and enumerates the three main forms such concurrency can take: strong, weak, and global management. Each form of concurrency is the subject of a separate experimental case study, conducted on a VAX-11/784 shared-memory multiprocessor, using the Mach operating system. The first study involves an IPC processor that can significantly improve the performance of communicating processes by exploiting strong concurrency. The second study looks at the scheduling of OS processing on a multiprocessor system, and shows how performance can be improved through policies that emphasize weak concurrency. The final study is an example of global management concurrency, in which the overhead of a complex real-time scheduling policy is significantly reduced by means of a dedicated scheduling processor. The use of carefully designed measurement tools and techniques makes the experimental results highly reliable. A major contribution of this work is that it distinguishes the benefits resulting solely from concurrency, and identifies the situations in which specialized hardware support is needed to achieve further improvements. Using these results, along with the cost estimates for alternative implementations, system designers can make more informed cost/performance tradeoff decisions. Consequently, this thesis contributes toward the development of a more soundly based hardware support design methodology.