User-level threads can exhibit excellent performance when they perform normal thread management operations, but when they meet kernel events such as I/O, page faults, and processor preemption, they can exhibit poor performance and even take incorrect scheduling decisions beside those of kernel threads. Scheduler activations model from Anderson et el., deals with this problem and we can get the performance of user-level threads and the functionality of kernel threads.
Windows NT supports the shared-memory multiprcessor system and takes the time-sharing scheduling policy. In this thesis, we propose a machanism that improves current parallelism supports of Windows NT. We implement the user-level thread system and the scheduler activations model in Windows NT. We describe the design decisions made, and the implementation details. To validate our implementation, we take performance measurements of the trivial parallel applications. We propose a new way of implementing scheduler activations model, which can be useful for the rapid prototyping to other operating systems that perfOrm the time-sharing scheduling.
