Please, next time use a sane quoting! The quoted text should be prepended with quoting characters, not the replies. On Wed, Oct 13, 2004 at 21:36:57 +0530, Nilanjan Roychowdhury wrote: > -----Original Message----- > From: Jan Hudec [mailto:bulb@xxxxxxxxxxxxxxxxxx] On Behalf Of Jan Hudec > Sent: Wednesday, October 13, 2004 6:42 PM > To: matrix reloaded > Cc: Anupam Kapoor; kernel; Nilanjan Roychowdhury > Subject: Re: thread awareness of the scheduler > > 1) Kernel can't replace kernel-threads with user-threads. Libc (since > libpthreads are part of it) can do it. > > >>> you mean to say that NPTL does this M on N mapping ??? I say, that NPTL is a user-land stuff. Thus the switch couldn't depend on kernel. But kernel can provide new API that makes the features possible. It does so. As I found from brief search on the internet, there is *NO* M-on-N mapping going on! The NTPL seems to be mapping threads to processes 1-on-1! The difference is, that clone(2) was extended to allow all these processes to have the same pid and thus show up as a single entry in /proc. Since /proc is what ps uses to show the processes, it shows the threaded application as a single process. > 2) User-threads are mostly useless. When one thread blocks, the whole > process does. And it can't be run in parallel on SMP. > >>> r u sure this happens in 2.4.20 + kernels ??? it looks like in > 2.4.18 the things are different. Here I was saying that I don't believe NTPL uses user-level threads. No, it does not use them. The difference that came with kernel is, that "thread groups" (that show up as a signle process) and some other API to simplify posix thread library were implemented. NTPL then took advantage of these new features. > 3) There is another thing kernel can do -- it can change semantics of > /proc to be by-pid and not by-tid. While each thread has different > tid, it has the same pid (pthreads want it). If ps does not see tids, > it can't show them... > > >>> This is what NPTL doc says...now in /proc it is PID not TID. Yes. And it is the *SOLE* reason for the difference in ps output. Well, in fact, the manager thread is also gone and some ABI is being used so it's all faster. But the basic principle -- that each thread is a process -- has not changed. ------------------------------------------------------------------------------- Jan 'Bulb' Hudec <bulb@xxxxxx>
Attachment:
signature.asc
Description: Digital signature
