On Sun, 4 Feb 2007, Jakub Jelinek wrote: > On Sun, Feb 04, 2007 at 03:12:32PM -0500, Bill Davidsen wrote: > > Arjan van de Ven wrote: > > >>Because user threading can avoid context switches, there will always be > > >>cases where it will outperform o/s threads for hardware reasons. > > > > > >actually.. switching from one "real" thread to another in Linux is not > > >an actual context switch in the hardware sense... at least this part of > > >your argument seems to be incorrect ;) > > > > > How does that work? Switching between kernel threads requires going into > > the kernel, user level thread switches are all done in user mode. > > > > Do you have some way to change o/s threads w/o going into the kernel? > > But going into kernel is not very expensive on Linux. > > On the other side, the overhead you need to add for every single syscall > that might block for the M:N threads and the associated complications > which make it far harder to conform to POSIX IMHO far outweight the costs > of going into the kernel for a context switch.
Agreed, definitely. A libpcl (using swapcontext(3)) cobench is about 50 times faster than an context switch measured by lmbench (although I have serious doubts about about the ability of lat_ctx to measure it - but that's another story) on an Opteron 254. One may say "Wow! Really?!?". The point is, who cares. We are talking about differences between super-fast (~2us) and ultra-fast (~0.04us). The time (and code) that you'll have to drop in the syscall path to handle M:N is very likely to make you lose way more of what you gain by avoiding an OS context switch (a "soft" context switch you still have to do it). Either use N:N (requires locking, but spread over multiple CPUs) or 1:N (I/O driven state machines or coroutines - no locking, once-CPU bound). - Davide - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/