Dear CFD's expert:

I have multi block RANS program in FORTRAN that runs on SGI Origin 3800. I did run the same program on PC single processor (Pentium III 1Ghz, 256 RAM) using Digital Fortran compiler. The result is the SGI is twice faster than the PC. Do you think if I use dual processors (e.g. Athlon MP: 2-AMD 1.2Ghz CPU'S) is going to double the performance (in terms of CPU time)? Does the compiler have to be taking into account for dual processor? Is Lahey or DVF Compiler also compatible for multi processors?

Thanks you for sharing the information.

Prahoro Nurtjahyo

From what I have been told, the compiler has to be designed for a dual processor environment, otherwise you will only see a 20% speed up going from a single 1.0 GHz to dual 1.2 GHz processors. Of course, if you run other app's simultaneously, you will see better overall performance.

Even if the compiler is designed to use dual processors, do not expect a 2*1.2=2.4 performance increase compared to a 1.0 GHz single processor. Probably more like a 1.5*1.2=1.8 speed up due to the shared memory bus.

If the code is well threaded, or MPIed, you might obtain a three times speed up with the duo Athlon. Experience with CFD codes in our institute shows that for identical clock speed, the Athlon is about 30% faster than the PIII. Plus, it is not that hard to have almost perfect parallelization (close to 100% efficient) in a duo CPU environment.. So you'd be looking at 2*1.3*1.2=3.12 times faster than the single PIII. Depending on your code, compiler, OS, etc, your mileage may of course not be as good.

It has also been my experience that the Athlon is about 30% faster, clockspeed for clockspeed, than a Pentium III. It also compares very favourably to a Pentium 4. We found that a 1.2 GHz Athlon was about 5% slower than a 1.4 GHz Pentium 4 (and half the price, for both CPU and RAM) I have now tested a dual Athlon 1.2 MP, and it is significantly faster on a single (non-parallel) job than a single CPU Athlon 1.2 GHz. Interestingly, doing two jobs at the same time seems to slow it down by about 30% on each job, although you still get the added productivity of two data points rather than one. The "double job" slowdown that we observed on a dual Pentium III was less marked (about 15%). The AMD 760MP chipset has a rather interesting memory bus design, which is probably the reason for this.

>> Interestingly, doing two jobs at the same time seems to slow it down by about 30% on each job, although you still get the added productivity of two data points rather than one. The "double job" slowdown that we observed on a dual Pentium III was less marked (about 15%).

The reason for the 15%-30% slowdown might be coming from the operating system you are using. Under linux, when running two jobs simultaneously on a duo penthium II box, each one runs exactly at the speed it runs on a uni-processor system. The same is true when running linux on a 4-cpu Alpha system: 4 jobs can run at once, each one not interfering with the others. I haven't yet measured the performance of a duo Athlon, but I would expect the same trend: there shouldn't be a slowdown caused by multiple CPUs coexisting on the same motherboard.

Mmm, that is interesting. The dual Athlon MP we tried WAS running Windoze 2000. Nevertheless, I have found on a dual PIII/800 running Linux that a single job runs quicker on its own compared to when a second job is also being run. However, (and we checked this by unplugging the one CPU) it also runs more quickly than on a single CPU computer using the same CPU! This is only partly due to the "idle" CPU (not strictly true, when running SMP without affinity the job is shared equally between the two CPU's) taking over system tasks, when you time it on CPU seconds it is still faster. I believe there is slightly more memory bandwidth available, and perhaps also some help from the cache on the second CPU?