[gmx-users] Re: Benchmarking Gromacs 3.2.1 on Power4

[Fiona Reid]

> Dear GMX-users,
> 
> I run benchmarks on IBM Power4 (aix5.2, 1.1GHz) using mdrun of Gromacs 
> versions 3.1.4 and 3.2.1 and single precision. The results seem a bit 
> disappointing as mdrun in version 3.1.4 was faster:
> 
>  		3.1.4		3.2.1
>  		-----		-----
> d.villin	4015.44		3645.888
> d.poly.ch2	2368.152	2155.152
> d.lzm.cutoff	1165.992	1154.328
> d.lzm.pme	760.728		747.456
> d.dppc		91.08		84.408
> 
> The same phenomena was observed in parallel results of d.dppc. I suspect 
> that mdrun vs. 3.2.1 shouldn't be slower, but have at least the same 
> performance than mdrun 3.1.4
> 
> Installation procedure of version 3.2.1 was carried out using the helpful 
> tips from Fiona Reid: 
> http://www.gromacs.org/pipermail/gmx-users/2004-July/011630.html
> The necessary inclusion of --disable-vectorized-sqrt option might (partly) 
> cause the slower performance of 3.2.1 mdrun. Has anyone else experienced 
> this phenomena of a slower mdrun in vs. 3.2.1 on Power4 and found a way to 
> overcome it?

See Erik Lindahl's message with the bug fix for Power4
http://www.gromacs.org/pipermail/gmx-users/2004-August/011655.html

I implememted this and found that for the DPPC benchmark Gromacs 
3.2.1 was 1-2% faster than 3.1.4.

Hope this helps. 
Fiona