用户空间中实现的线程库可以支持超线程吗？

Question

Assume a multiprocessor architecture with an OS which may or may not support kernel-level threads

Correct me where I go wrong:

• If a thread library is implemented entirely in user space, then the management of threads is done in user space (creation, thread tables, stack info, ect).
• Thus, even though a process may have multiple user threads, the kernel only sees 1, single-kernel-threaded process.
• Thus, the kernel schedular assigns CPU usage time to the process as a whole; the user-space thread library is responsible for time-slicing this CPU time amongst it's user threads.
• ( corollary 1 ) process A with 20 user-threads will get same priority as process B with 1 user-thread, and so the threads in process A get approximately 1/20th CPU time as thread in process B
• ( corollary 2 ) user threads within the same process will never be hyper-threaded (ie 2 threads executing on separate CPU's at the same time)

Answer 1

Your first three assumptions are correct.

Corollary 1 depends on the OS scheduler. Scheduling can be based on threads only, not processes so there is no inherent guarantee that processes with different thread counts get the same total time.

Many user-space schedulers go a hybrid route and schedule m user space threads to n OS threads (with m >> n ), thus avoiding some of the overheads of OS thread creation. There is no way of magically achieving concurrency without resorting to OS mechanism to bootstrap it.