Hi,

Have anyone measured the parallel speed up using multiple PC's with CFX 5 ?

Either local (multiprocessor computers or distrubuted ) ?

Any comments ?

Thanks in advance.

Kind Regards

Roued

With a large enough grid:

- 2 processor machines running NT/2000 get something like 1.5-1.7.

- 2 processor machines running Linux get 1.7-1.8.

- Running on a cluster of NT or Linux machines would scale from there. It's typical to see parallel efficiency on the order of 70-80%.

Note that scaling is very problem size dependent, i.e., you need enough grid to keep the matrix assembly and linear solver busy. A good rule of thumb is to have something like 30,000-40,000 nodes per partition.

I also want to add that you will fare better on a decent unix workstation (speedups of 1.9-2.1 on dual cpu boxes are typical) because the cache and memory subsystems, as well as dynamic process managment, are so much better than a PC.

Dan.

Hi,

Have you compared the performance of Unix (non PC's) to PC's ?

Or P3 to P4 ?

It sounds like it.

Kind regards

Roued

I can't say anything about a pentium 4, as I've never run CFX-5 on one. But here are some numbers, using a 800MHz PIII as a reference:

PIII 1GHz 1.02

Alpha 833 MHz 2.05

PA8600/552 MHz 1.48

R12000/270 MHz 0.71

R10000/250 MHz 0.60

RS6000/43P/260 0.85

These numbers were obtained for a serial run on a 160,000 node hex grid.

Dan.

I haven't used CFX5, but for our in-house codes and for two other commercial codes that I have tried the 1.7GHz P4 is about twice as fast as the 1GHz PIII, provided the P4 has dual-channel PC800 RDRAM. This speed difference is not seen in normal application, but has been verified by several others for CFD applications.

The P4 is also faster than the 833Mhz Alpha chip, although the latest 1GHz Alpha chip outruns it slightly I think. So you should definitely check out the P4 - we are about to replace our high-end HP desktops with dual CPU P4s running Linux.

Yeah, I got exactly what you said, 1.45 times the serial run, in a simple geometry with 60k cells on a Dual 850MHz NT system.


cfd guy

Hello

All this parallel talk is very interesting! I have a few questions as I am a complete novice

Scalability/Scalable - I have come accross this new buzz word quite often, what is it in reference to i.e. what does the term mean in computer hardware terms?

These figures that people are quoting, is it an index or something? How is this figure determined?

What exactly is parallel efficiency?

Does anyone have any experience of the new Intel Itanium processor yet?

Has anyone got figure on the Zeon processor?

It Would be great to get some answers guys/gals

Cheers

kwabs