The continuing history of Linux threading is IMO an interesting one, and should answer your question, so here's a potted history so far.
The original linux
thread implementation, called LinuxThreads, was built on a model where each thread was a process in its own right, with a manager thread to control creation and destruction of "threads". However, it didn't scale well, there were bottlenecks, some posix thread features could not be implemented, and the overheads were high. Each thread had its own PID.
Your linux system most likely uses LinuxThreads.
In the rush to improve threading on Linux, the state of which has been one of the main obstacles to Linux being adopted on an Enterprise scale, two main projects worked in parallel (and in competition) hoping to be adopted as the Linux threading model. They are the NGPT project and the NPTL project.
In late 2.4 kernels the NGPT (next generation posix threads) work can be back-ported to the kernel. In this threading model the actual user threads (eg. threads that you spawn in your
Java proggie) are different from the kernel threads and scheduled separately. This is generally called the M:N threading model, where there are M threads in the process, and N underlying kernel threads to handle the actual processing. However, this had problems too, with thread priorities, scheduling, and again some posix features could not be implemented.
The threading model in 2.6 kernels is based on the work of the NPTL (Native POSIX Thread Library) project, which implements a 1:1 threading model, ie. one kernel thread for one user thread. This should mean that for each thread that you spawn in your java proggie there will be a kernel thread spawned. The kernel's internal threading infrastructure has also been rewritten to allow the Native POSIX Thread Library (pthreads) to run on top of it (but since Java threads don't exactly match POSIX threads if your interest is in java thread programming this may be irrelevant for you).
It seems that the NPTL project won the day and has secured its place in Linux's future. Hence its appearance in the 2.6 kernel.
Where threading is concerned, the kernel can move each kernel thread from one processor to another in an SMP (symmetric multiprocessing) configuration. This is not possible with independent user threads. The important difference between Linux kernels 2.4 and 2.6 is that with a 1:1 threading model (2.6 kernels)
you should get far better performance given a multi-processor machine or a machine with a processor that implements hyperthreading (eg. the Intel Xeon).
This has been good news for Enterprise-class Linux.
HTH.
[ July 08, 2004: Message edited by: George Brown ]
