IBM 610 workstation computer 3466

+--------------- You may be tunnel visioning; memory isn't the only wire that has to be going to all CPUs; you also have to think about devices that pick up and deliver little bits. +---------------

(*Sheesh!*) I've already said it twice -- this is the THIRD TIME, Barb! On Opterons & Origin-Altix, device DMA *is* automatically cache-coherent, in hardware!!

+--------------- I'm NOT talking about user programs; I'm talking about the operating system that is always resident...that piece of code. +---------------

Well, Irix & Linux & all of the BSD's *are* also "SMP" in the operating system, for almost everything. System calls can run on whatever CPU the user was on when the call was made -- none of that silly "master-slave" stuff the early dual-CPU KA10's were stuck with. Yes, for best performance, *some* things might be deliberately "pinned" to one CPU (e.g., one might "pin" interrupts from certain devices to certain CPUs, for various performance reasons), and of course, the scheduler & memory allocators themselves need to know where the CPUs & memories are and how many-much there are, but on the whole modern operating systems are just as CPU-insensitive as user code.

+--------------- I don't like that number 4. This is ridiculous. It should have been in the thousands by now; it's been 25 years! +---------------

It is almost thousands for SGI's Origin & Altix. Give the Opterons a couple more year; they're still babies, by comparison. Except... they're *fast*...

+--------------- This sounds more like multi-CPU rather than an SMP setup. IOW, one CPU buttigns computing tasks to the others in the schedular. Or, perhaps, a few CPUs buttign tasks...nope, can't happen. This would require scheduling hand shaking. +---------------

Well, DUH! The "hand shaking" needed is called a "lock", and lock overhead *is* a major component of any high-CPU-count system. In modern operating systems (Irix, Linux, loveBSD) the scheduler runs on *all* CPUs, and whenever any CPU needs something to run it jumps into the scheduler, finds something that's optimal (or at least "good") to run on that CPU, and runs it. Modern schedulers try to keep processes running on the same CPU across multiple schedulings, when then can, to avoid excssive cache thrashing. This is called "cache-affinity scheduling", and is old technology.

+--------------- SMP setups have each CPU choose which computing it's going to do from a grocery list. +---------------

That's what I've been trying to tell you! *All* modern operating systems do this by default. The operating system itself is a multi-CPU, multi-threaded, cooperative computation that treats the CPUs, memory, and I-O devices as a huge pool of shared resources.

+--------------- TOPS-10 called this list the runtime queue (that's not quite right). +---------------

They're called "the run queue(s)", and yes, there can be either just one or several of them, depending on the specific operating system and-or the system configuration. The queue(s) have to be locked for the small time one is adding or deleting items to-from them though there are a clbutt of "lock-free queues" that can be used in some situations.

-Rob

----- San Mateo, CA 94403 (650)572-2607