On Fri, 25 Oct 1996, Matthew Jacob wrote:

> 
> 
> 
> Yep. Insanely so. To the point of using LDSTUB (atomic load/store
> unsigned byte) in machine dependent code on platforms that can
> *never* be MP (sun4c). Atomic instructions cause a flush of
> write buffers and a halt of SBus activity.

This does seem silly, though I would guess there is some extra 
development and test overhead in doing this stuff two different ways.  
I'm wondering how much longer sun will support the 4c line.  If it goes 
beyond 2.6 I'll be surprised.

> 
> I really don't think that Solaris could be taken as a sane example
> of what "true" multiprocessing should be. A much better example would
> be sorta like early SGI variant kernels I saw at Kubota (Ardent):
> one atomic operation (spinlock, with h/w timeout/death) and a P/V semaphore
> built on top of that to replace/supplement sleep/wakeup. Period.

Hmm, does this approach make sense in the context of systems with say N
or more CPU's?  The MP performance of Solaris has been getting better 
and better each release, or so sun says.  If you have say 32 CPU's, 
can you get away with doing things any less stringently and still get
the same level of performance?

It seems like sun may have taken an approach that hurts the low end
in order to push the performance on the high end.

I've been piecing together a sun 4/600 system.  I've got two of the 
old ROSS 40MHz duel proc boards.  Once I get enough RAM for the system
I'll be bringing it up.  Anyone got any suggestions for how to measure
performance?