slow runtimes with intel fortran

[jensenb01]

Hello.

I have an intel fortran run time problem. I have written fortran 77 sourcecode for Goddard Space Flight Center. I wrote and debugged the code using Compaq Visual Fortran. Goddard wanted to run the fortran on their SGI, Linux, Itanium. Goddard purchased intel fortran to do this. Unfortunately, the SGI runtimes were always much slower than runtimes using Compaq fortran on a regular PC. I tried compiling the SGI (intel) fortran with many different options, trying to optimize the compilation to run faster. I have not been able to discover the source of the problem. The work around has been installing Compaq fortran on the PCs of those who need to use the fortran. However, we would like to use the intel fortran if we can solve the runtime problem. I just made a run on my PC and also on the linux SGI:

Runtime using my regular DELL PC and Compaq fortran: 19 minutes
Runtime on the SGI linux itanium with intel fortran: 71 minutes

The SGI is a high end computer and normally runs much faster than a regular PC. The majority of usage on this computer is running the NASTRAN structural finite element code. NASTRAN has a lot of fortran in it and uses intel fortran as well.

Brent Jensen
Phone: 301-286-5906
Email: bjensen@mscmail.gsfc.nasa.gov

[jimdempseyatthecove]

Your Itanium version might be compiled with runtime checks enabled? (Array Bounds, Uninitialized Variable testing, etc...) whereas the PC version is not. Also check for vectorization.

Jim Dempsey

[TimP]

MSC-NASTRAN spends most of its time in asm code optimized specifically for IA-64, so that's more of a counter-example to the implicit idea that ifort would perform well on IA-64 without attention to detail.
If you want help, you should profile to determine where the excessive time is spent, make sure you used appropriate compiler options, then show us the opt-report for that subroutine (and, preferably, the source code).

[Steven_L_Intel1]

Vectorization is not used on Itanium.

[TimP]

The nearest equivalent, software pipelining, is partly enabled at -O2 (default, when -g is not set), and fully on at -O3. A wild guess is you need such optimizations, and should look to see why you don't have them.