Hierarchical schedulers, performance guarantees, and resource management

An attractive approach to scheduling applications with diverse CPU scheduling requirements is to use different schedulers for different applications. For example: real-time schedulers allow applications to perform computations before deadlines, time-sharing schedulers provide high throughput for compute-bound processes and fast response time for interactive applications, and gang schedulers and cluster coschedulers permit tightly-coupled parallel applications to achieve high performance in the presence of multiprogramming. Furthermore, individual members of these broad classes of algorithms make tradeoffs that may or may not be appropriate for a given situation. In order to take advantage of these diverse algorithms, we permit schedulers to be arranged in a hierarchy - a root scheduler gives CPU time to the schedulers below it in the hierarchy and so on until an application thread is scheduled by a leaf scheduler. This architecture has a number of advantages.