Return to BSD News archive
Path: sserve!newshost.anu.edu.au!harbinger.cc.monash.edu.au!bunyip.cc.uq.oz.au!munnari.oz.au!constellation!news.uoknor.edu!ns1.nodak.edu!netnews.nwnet.net!oracle.pnl.gov!osi-east2.es.net!lll-winken.llnl.gov!koriel!male.EBay.Sun.COM!engnews1.Eng.Sun.COM!engnews2.Eng.Sun.COM!orpheus!jackv From: jackv@orpheus.Eng.Sun.COM (Jack F. Vogel) Newsgroups: comp.os.386bsd.misc Subject: Re: dual procesor motherboards the way forward? Date: 10 Oct 1994 19:15:49 GMT Organization: Sun Lines: 72 Message-ID: <37c3t5$a4m@engnews2.Eng.Sun.COM> References: <Cww6x9.1A0@gnome.co.uk> <tgray-0710940002460001@192.187.228.31> NNTP-Posting-Host: orpheus.eng.sun.com In article <tgray-0710940002460001@192.187.228.31> tgray@netcom.com (Tom Gray) writes: >Wouldn't it would be great if someone produced an inexpensive >mass-merchandised motherboard with between six and eight 40MHz 386DX clone >chips installed? With the right (SMP aware) operating system, a $500 >motherboard like that would scream. Won't happen, mainly because in order to do MP based on a 386 requires a bunch of off-chip support on the CPU board (or motherboard). So when you do it you face STANDARDS issues. This is why its only with Pentium (and only the newer chips at that) that there is a potential for a viable MP market. The Pentium has things like 'cpuid' identification, bus arbitration, and so forth IN THE CHIP. I am not an expert in X86 architecture, but I believe it was also ONLY as of the 486 that this architecture had an ATOMIC 'compare and swap' (which they call compare and exchange :-), something that is desireable in implementing MP locking. Don't get me wrong, 386/486 MP HAS been done, its just that its based on complete systems from certain vendors, and the OS support has to be written to that SPECIFIC hardware. Now, with Pentium and the Intel MP spec, maybe there can be a 'clone' "mass-market" as you say. >No, Windows NT supports full SMP (Symetrical Multi-Processing). Not only >are the applications automatically scheduled across available processors, >but so is the operating system itself. Further, the applications >themselves are broken down into threads which are split accross >processors. The great thing about all of this is that it is really very >inexpensive (<$400). Ya, so... You can put some super-duper turbo engine into a Yugo, but the question would still be, WOULD YOU WANT TO BE SEEN DRIVING IT :-) ;-). [ previous poster asks about a master-slave CPU arrangement...] >This is called ASymetrical Multi-Processing (ASMP) and, although it much >easier to implement and program for, it does not offer as much performance >as SMP for non-compute intensive applications. This is because most >typical applications spend most of their time in the O/S (doing windowing, >networking, I/O, memory management, etc..), which makes the "root" >processor a resource bottleneck. Still, for certain types of applications >(simulations, etc..) it does a very good job. Keep in mind that symmetry/asymmetry is not a binary choice, it is a spectrum with many shades of gray inbetween. I prefer to think in terms of how coarse/fine grained the implementation is. SunOS 4.X is fairly course, it has a single global lock, STILL the kernel CAN run on any CPU so it is NOT really asymmetric. Also, there ARE some things in the kernel that can be done outside the lock AND it can run user-land stuff in parallel. So, its still a win. It is FAR easier to make an existing UP kernel MP via a course grained approach, and then SLOWLY move things out of the main lock, than try to fully thread the thing up front. MP locking and debugging can be a NIGHTMARE :-) :-). You have NO IDEA until you have actually worked on such a beast, old notions of 'safe' coding practice get TOTALLY blown away :-). >>Anyway I am sure the FreeBSD and NetBSD teams >> could have some fun with this after they have tired with porting 4.4 >> BSD. > >It would sure distinguish *BSD from Linux. Hmmm, ya, well we shall see. Cheers, -- Jack F. Vogel Sun Microsystems Inc. jackv@Sun.COM