...Does NT support both user level threads and kernel threads? I know that Solaris supports both, and when you invoke Solaris thread library, and thread library calls gets translated into system calls and are handled inside the kernel and they have special hardware to handle context switch. I am not sure how NT implements it
Byron, I'm not sure about this because I read about it a long time ago. But from what I remember, the difference between NT and Solaris when it comes to threads is that the thread is the kernel scheduling entity in NT; that is, the kernel knows about all the threads in a user process and the kernel schedules them. I think that Solaris instead has this concept of "lightweight processes" that are scheduled by the kernel, but don't correspond one-to-one with user threads. The Solaris user-level thread library schedules threads by multiplexing them on these "lightweight processes", usually without entering the kernel.
It's another reason NT doesn't scale as well as Solaris. You can have millions of threads alive in a Solaris machine and the scheduling load on the kernel stays reasonable. Because: the kernel doesn't need to know about all the threads at the user level. But with NT, the more threads you have, the heavier the scheduling burden on the kernel, since it knows about all of them.
They way I heard it explained, it's like if in New York City everybody had to have their own car, that would be NT threads, and the more people you had, the more cars and the more gridlock. With Solaris threads, everybody takes a "taxi" (the lightweight process), so the gridlock goes up much more slowly with more people, since a taxi services many people and you add taxis instead of cars.
This doesn't mean that NT would lose if you ran a benchmark and just measured context switch time with a tiny number of threads. It might win. But wait till you try to run NT on 64 processors. Answer: you don't.
Maybe that doesn't make any sense but oh well. It's another reason why NT is a dog. I think M$ has all kinds of people like Tandem and Sequent etc. trying to fix this problem since of course they can't do it themselves. Who knows, they may succeed one day.
Regards,
--QwikSand |
