From: ahemavathy (ahemavathy_at_hotmail.com)
Date: Sun Jan 25 2004 - 19:22:50 PST
To achieve high performance parallel computing, user threads should be aware of the kernel events and also have complete control over the threads running on the processors allocated to it. This is what the paper is about..how to give that control and awareness. User threads built on top of kernel threads suffer from the problems of heavy kernel threads although user threads are themselves lightweight and flexible. Although user threads could work around the problems, the kernel does not provide adequate support for user threads to do that.
This paper is about the new kernel interface and user level thread system that provide the flexibility to user threads. There is communication between them in terms of scheduler activations. Basically the kernel informs the user level when an additional processor is added, processor is preempted, scheduler activation is blocked or unblocked. Similarly the user level informs the kernel when it needs more processors or when it has free processors. To improve performance, discarded activation schedulers were collected for reuse. Such a system was built by modifying Topaz and FastThreads.
It was interesting to note that although the individual thread and upcall performance is not good, the overall application performance is very good especially when compared with the old scenario of user threads running over kernel threads where when a user-level thread is blocked in kernel, the kernel thread is also blocked. Overall this was an interesting paper which made me wonder how current systems handle threads.
This archive was generated by hypermail 2.1.6 : Sun Jan 25 2004 - 19:22:58 PST