[Beowulf] x86-64 NUMA vs SMP kernel: appl. performance?

[Greg Lindahl]

It looks like Bill covered most of the questions, but I'll clean up a
loose end:

> I have not yet done a similar comparison against the 3.4 GHz nacoma and
> intel's 8.1 compiler.  I'm not sure if the 8.1 compiler will cripple
> x86-64 running like the previous version did.  Nor have I tested pathscale
> 1.3.

The 8.1 Intel compiler doesn't cripple EM64T, but it has the same
scheme for crippling AMD64 -- if you use the best optimization flag,
it inserts a runtime check for a "Genuine Intel" cpu id. If you edit
out that check, our preliminary results are that we beat them on AMD64
today, although they did teach us a few things. That's no surprise, as
the microarchitecture is considerably different between the EM64T and
AMD64 implementations, and Intel hasn't tuned for AMD64's
microarctitecture at all.

The fun part is guessing what our performance is going to be when
we're through tuning our EM64T optimizations. Our 1.4 compiler release
(early October?) will support EM64T, with full tuning in the next
release (which will also be our OpenMP beta.)

-- greg