From: "Lusercop" <`the.lusercop'@lusercop.net> > On Thu, Jan 30, 2003 at 01:18:41PM +0000, Jonathan Peterson wrote: > > Anyone using Linux on anything with lots of CPU's? Attempt to do > > searches for 'linux smp' on google tends to get me documents last > > updated in 1997. > > I used to run it happily on 2 CPU's, but I wondered if anyone was doing > > it seriously on more than 4.
Most of the stuff on the web suggests that linux doesn't (or didn't) scale particularly well past 4 processors on Intel and I doubt there are many people in the world running this since Solaris on UltraSPARC is more common in this niche. Although I think a lot of current linux development work is on SMP and multi processor scalability so if you were brave enough to run a bleeding edge kernel with dodgy patches you might get better results. And as ever in the linux world what I saying may be totally out of date as of last week. 2.6 is likely to have NUMA. > I could be wrong on this, and I suspect that doop (Jon Chin) or quidity > (Alex Gough) are better people to ask, but basically over about 4 processors > SMP is a bad idea, and you start to want an architecture that is MPP-based, > where you have effectively independent computers connected by a very fast > network-like bus. I'm not sure what the big machines (E10k/E15k/S390/etc) > use, but I would have thought that it's unlikely to be SMP, just due to > the need to occasionally lock the bus, which scales appallingly badly. I suspect MPP (message passing) versus SMP (shared memory) is a CS religious issue and don't claim to understand all the issues but isn't MPP generally very slow? I suppose you have to draw tradeoff graphs. AFAIK the big Suns *do* use SMP and I can't think of any mainstream UNIX server kernel which uses MPP (maybe Alphas running OSF?). Although Apple's Darwin is based on Mach (and other things) there seems to be a bit of doubt about how well supported the Mach stuff is and they seem to use SMP for their multiprocessor support. What's better in theory often doesn't happen in practice since if the CS academics of a decade back had predicted future technology correctly then we would be all running microkernels on RISC chips today. And we aren't. -- 1024/D9C69DF9 Steve Mynott <[EMAIL PROTECTED]>
