[gmx-users] optimization of gromacs 4.0.5 on AIX Power6

[Mark Abraham]

Silvia Giuliani wrote:
> Dear All
> 
> Has anybody tried version 4.0.5 on an IBM Power6 (4.7Ghz) cluster?
> We have installed the program with the following configure options 
> (using fftw we installed separately):
> 
> ./configure --with-fft=fftw2 --enable-mpi --enable-float CC=mpcc_r 
> F77=mpxlf_r CXX=mpCC_r CFLAGS="-O3 -qarch=pwr6 -qtune=pwr6 -qcpluscmt"
> 
> It compiles and works but the performance is poor compared to our 
> Opteron  Cluster (2.6Ghz)
> For example with 8 procs for a small protein in water we get 5.8ns/+day 
> compared to 9.6 ns/day on the Linux cluster, which is a bit disappointing.
> Since there alot of compiler options to play with we would be grateful 
> if somebody who has already found a good set would share them with us.

If this processor does not support PowerPC Altivec, then the reason is 
straightforward. GROMACS gets its speed from optimized non-bonded 
interaction kernels, and your machine doesn't have any written for it. 
You can see in the .log file whether generic or optimized kernels are 
being used. Optimized kernels exist for x86s, so they're fast.

Given the above assumptions, your GROMACS installation will be using 
generic kernels written in C. You should certainly try the 
--enable-fortran configure flag to see if the equivalent in FORTRAN is 
better optimized by the compiler.

You can try -O4 or -O5, but these tend to have little utility for GROMACS.

Mark