> > This is almost the perfect application for clusters. Clusters don't have
> > the tight communcation network of SMPs or ccNUMA but tend to be quite
> > cost efficient which is a good fit for the typical raytracing app.
> Clusters also don't suffer from the bus clogs of shared-memory machines.
It's however easy to comeup with communication patterns that will make a
cluster perform worst case. Basically it all condenses into graph theory,
can you represent a communication pattern efficiently in the communication
network given by a certain architecture?
> I'm curious: are there any (cheap) systems out there with some shared
> and some unshared memory, and multiple buses? This would seem to allow
> a wonderful combo of the speed of local exclusive memory with shared
> memory faster than any network.
Think of the Origins as something like that; they're kind of clusters of
2 CPU SMPs. Each processor has access to all memory in the entire system.
If necessary the VM system could replicate pages in processor local
memory as kind of a level 3 cache thereby eleminating the shared memory
overhead. No idea if IRIX does that, but that at least an obvious trick.
In any case the machine's architecture makes representing many communcation
patters efficiently in software way easier.
HP's big HPPA systems come to mind as well, maybe somebody knows more about
their stuffs ...
> Or is this perhaps what they call a supercomputer?
It's indeed what they call supercomputer even though good ol' Seymour
probably would spew out his soul for using this word on a CMOS thingy with
caches ...
Ralf
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/