[gmx-users] Re: latest benchmarks (Itanium/Xeon/Opteron)

[Michel Cuendet]

Hi,

>>>>>   The --disable-software-sqrt option doesn't help. I noted also that
>>>>> gromacs single precision computes wrong results (box exploding
>>>>> systematically), whereas double precision doesn't take more time and
>>>>> gives more reasonable output.
>>>
>>> This shouldn't be... What system is that? Is it reproducible, i.e. the
>>> same tpr gives different results in single and double?
>>
My system is a big protein complex, which I'm trying to pull apart to 
measure an affinity. Gromos force field, PME, Nose-Hoover. (Explosions 
in single precisions also occur when no pulling is done).
  Yes it is reproductible, because I ran 64 Jarzynski trajectories in 
parallel, which all crashed after a few Ksteps. Running the same 64 
trajectories in double precision has been stable for 3 weeks. Also, 
copying the same .tpr to an alpha machine or an athlon produced behaved 
results in single precision

>I think I've disabled software sqrt by default on ia64, so that  
>explains the lack of effect.
>
In fact, I tried --enable-software-sqrt also, which has no effect either.

>
>
>To help isolate the problem, check if it occurs:
>
>(A) when you use gcc
>
It doesn't work. Compiles fine, but crashes at runtime because of a PME 
problem.

>
>(B) when you use a lower level of optimization (e.g. -O1)
>
I'm gonna try this out. Tomorrow I will meet an HP guy who was 
incredibly enthousiastic about trying to optimize gromacs on his 
itaniums. Let's see what comes out of that ;-)

>
>
>Using the Intel Fortran compiler might give you slightly better  
>performance, otherwise you'll have to wait a *little* while longer for  
>the assembly loops.
>
Option --enable-fortran does not compile. Problem with linking.
Oh, please post these assembly loops as soon as they are ready, can't 
wait !!!


Thanks,

Michel