Hierarchical Schedulers in the PM System-Architecture

Today’s computer and network architectures provide means for parallel execution of computing intensive applications. The primary operating system abstraction for supporting parallelization of applications is the thread. However, current implementations have severe deficiencies: kernel threads have too much overhead, and user-level threads are not integrated well enough into the operating system. Furthermore, an application may gain an unfair advantage from the scheduler by forking many threads. In this paper, we propose a hierarchically structured scheduling system. With a homogeneous integration into an open operating system architecture, the disadvantages of other thread concepts are avoided. In addition, a distribution of computing time among all applications and application subsystems is easily possible, independent of the number of threads.