Question about the PASCAL compiler

[susonic]

Hi all,

Recently I had bought a new system which is based on AMD cpu. And I was wondering how fast it is.
I had run under same condition such as same OS(SUSE 10.3), PGI compiler, same NETCDF version, same MPI, same SVN ROMS. But the new version of system did not show good performance. Therefore I had sent an Email and got an answer that
'Since the customer used none optimized compiler with the application for Opteron CPU and cause low performance of AMD platform'

Hence, I'm trying to use the AMD based compiler named 'PASCAL'. Is there anyone who had used 'PASCAL' and if so, would you show me the script?


Below are the test cases and comparison

model a. intel Xeon 5430 Harper town (Quad core) x2 => (8core) 8gb memory
model b. intel core 2 quad Q 6600 x1 => (4core) 4gb memory
model c. AMD Opteron Barocelona (Quad core)x2 => (8core) 8gb memory

test case : ROMS Benchmark case 1, 2, 3
case 1
model a 1min 30sec
model b 3min 30sec
model c 3min 3sec
case 2
model a 7min 58sec
model b 16min
model c 14min
case 3
model a 48min
model b Not Available(maybe memory)
model c 1 hour 25min

If there is any comment about this please let me know.

Thank you.

-Peter

[kate]

I have not heard of a Fortran compiler named "PASCAL". PASCAL is/was a computer language which used to be used for teaching - it was the first one I learned. I've never heard that it had great performance, but that's moot given that we're not rewriting ROMS into PASCAL.

[susonic]

Thank you for the reply. I was able to find an information about PASCAL.
Yes like you said, it is not a fortran compiler.
Is there a way to improve the performance with this AMD based cpu?

Any suggestion would be appreciated.

Best,

-Peter

[shchepet]

Several years ago I have tested a FORTRAN compiler called PathScale and compared
it against the usual suspects -- ifort and pgi. What I found is that when using
intel CPUs (Xeon, Pentium 4 at that time) ifort was definitely superior to both
alternatives; when using Opteron CPUs (these were older CPUs, I tested Opteron 248,
dual-CPU, single core; and Opteron 848, quad CPU, single core), I found that all three
compilers are very similar in performance and do equally good job. Eventually I lost
interest in both PGI and PathScale because they are commercial products, while intel
in freely available.

PathScale compiler in also known as EKOPath, and also known as "every known optimization".

It is designed by the same people from company called MIPS who designed Fortran and
C compilers for SGI back the first half of 199x (until Cray Research was bought by
SGI and took over SGI compiler development in 1997: thus 6.x compilers, whic 7.x
were designed by Cray (in reality 7.0 was terrible; 7.1 was bad, and starting with
7.2 the Cray Res.-designed compilers were able to match that of MIPS. Back in
199x-95 MIPS complilers were the best ever in term of utilization of the frnction
of theoretical peak speed of the CPUs.

Once there people were laid off by SGI (Capitalism is ugly!), they started their own
company and produced PathScale compiler.


Regarding you performance results, you are getting results which are pretty consistent
with what most other people are getting: that $180 piece of silicon called Q6600
matches in performance processors which are fancier and more expensive, making it
by fart the best "bung for your buck" ever. Opterons are slower nowadays in terms of
per-core speed than contemporal Xeons.

[gcreager]

There are several Opteron optimization options in the Portland Group compilers (and you can get a 30-day eval copy for free). A quick google for 'gfortran opteron optimization' suggests something like this might work well for you:
-march=nocona -ffast-math -funroll-loops -O3 -msse3 -ftree-vectorize

Most compilers have optimization now for Intel and Opteron CPUs.

gerry

[susonic]

Thank you for the answer, Schepet. That's valuable information for me.
And thank you, Gcreager. That's another important information. I'm going to try that one.
I really hope that this new opteron would get better.

Best,

-Peter