Re: Multi-processor?

In <URL:news:local.netbsd.acorn32> on Wed 07 May, Ben Harris wrote:
> In article <[email protected]> you write:
> >I saw on Drobe yesterday, and in csa.announce today that Simtec have a
> >clearance sale (see http://www.simtec.co.uk/products/clearance/) inculding
> >Hydra multi-processor cards for ukp50 (+VAT and carriage).
> >
> >Can NetBSD/acorn32 make use of multi processors using the Hydra card?
> 
> No.  Well, if you use code from the bjh21-hydra branch, you might be able
> to spin up the slave processors and run some kernel code on them, but
> that's about it.

There are better ways to spend 50 quid on ARM kit then.

> > What would be needed to enable it?
> 
> A really rather large amount of work.  My work on it stalled when the
> inter-processor interrupts didn't seem to be behaving as the documentation
> said they should, then I moved house and I haven't actually reassembled the
> Hydra since then.  Other problems include:
> 
> * The lack of any coherency between the CPU caches means that you need a
> cache flush after every lock operation, and you need to rewrite the
> simple_lock routines not to rely on cache-coherency.  This also means that
> the finer-grained NetBSD's locking gets, the slower the Hydra will go.
> 
> * The new ARM pmap is explicitly documented as being fundamentally
> incompatible with SMP on low-vector CPUs.
> 
> * Much of the ARM MD code needs updating to understand SMP.
> 
> * Once you've done all that, you end up with something slower than an
> SA-110 anyway.

Is hydra capable of takin SA-110 processor cards or is it tied to ARM6/7
processors? How about Kinetic cards? I guess if it was worthwhile someone
would have made use of it and Simtec wouldn't heve the hydra cards in a
clearance sale.

> I'll probably have another attack at the problem at some point (if only to
> merge bjh21-hydra back into the trunk), but I don't expect anything very
> useful to come of it.

Thanks for the detailed response.


-- 
[email protected]