[molpro-user] problem with parallel molpro

[Andy May]

Carsten,

We know about some problems with MKL on systems with AVX instruction, 
but it doesn't look as though that instruction is present on your 
machine. Perhaps you can do a test just to be sure it's not related to 
MKL and set:

export MKL_CBWR="COMPATIBLE"

before running the job, and also try with the default MKL link options 
suggested by configure. There was a previous thread about MKL here:

http://www.molpro.net/pipermail/molpro-user/2013-May/005394.html

which might be of interest.

There is another more recent tarball of the code compared to the one you 
used, it contains some fixes which might be relevant concerning 
overlapping vectors being copied.

Is there any significance in the fact that the non-working output used a 
node named 'gpu01', whereas all other nodes in the outputs are named 
'nodeXXX'?

Best wishes,

Andy

On 03/07/13 10:54, "Dr. Carsten Müller" wrote:
> Dear Molpro experts,
>
> we are currently testing different parallel versions of molpro 2012 on
> this computer
>
> http://www.zedat.fu-berlin.de/Compute/Soroban
>
> and have difficulties to get consistent numbers. The outputs s0000_2.out
> and s0000_2.out_8 where running on three cores on three different nodes or
> 4 cores on one node, respectively. However, only the latter output gives
> an MP2 correlation energy which is consistent with older Molpro versions.
>
> In one attempt (s0000_2.out_4_correct-O0_2N), turning of all optimization
> of the compiler and running the job on two cores distributed over two
> nodes we managed to get the correct value for MP2.
>
> We did use the latest molpro version. Parameters for configuring the
> global arrays (mysetup), molpro (mysetup_impi-5.1.1) and the final CONFIG
> file are attached.
>
> Can anyone suggest where to look for the source of this error?
> Feel free to ask me for more details.
>
> Regards,
> Carsten Müller
>
>
>
> _______________________________________________
> Molpro-user mailing list
> Molpro-user at molpro.net
> http://www.molpro.net/mailman/listinfo/molpro-user
>