[molpro-user] MPI error: Reply to Christan Stemmle

[Peter Knowles]

Dear Christian,

We have unfortunately found it quite challenging to maintain FCIQMC embedded in Molpro, and can only apologise for the frustration you experience.  In fact, it is probably not worth putting a lot of effort into solving the problems you experience, since in the end you will get access only to a quite old version of FCIQMC that is probably significantly slower than the latest.

Going forward, we have solved these problems by making a robust interface between FCIQMC and Molpro that allows one to use in a flexible way the latest version  of FCIQMC, rather than having it bundled with Molpro.  We are expecting to make this available in the next release of Molpro in a few months’ time.

Best regards,
Peter Knowles 

> On 5 Apr 2018, at 02:54, Nike Dattani <dattani.nike at gmail.com> wrote:
> 
> Dear Christian,
> Your error looks to be exactly the same as the one described by Hans Hogreve in the link you sent. Since you have an error with the MPI Broadcasting, can you tell us what MPI library was used when you installed MOLPRO? FCIQMC is parallelized with MPI, so at the stage where you got this error, it's trying to send a copy of some data to all cores that you have available. This data delivery (MPI broadcast) is done by the MPI library that was used to build your version of MOLPRO. The installation manual says "Intel MPI, Bull MPI, MPICH2, and Open MPI have been tested", so if you can check that you're using one of these, and the versions are compatible with everything else that was used, you might find the source of your problem. 
> 
> Alternatively, if you turn off MPI and try on a single core of a single node you should not get this error, but the calculation will be approximately a factor of N slower where N is the number of cores you have available, because the speed-up when going through the loop over determinants in FCIQMC is (almost) linear in N, and you would be losing one of the main benefits of FCIQMC.
> 
> With best wishes!
> Nike
> 
> ----------------------------------------------------------------------
> 
> Message: 1
> Date: Tue, 3 Apr 2018 10:37:14 +0200
> From: Christan Stemmle <christian.stemmle at fu-berlin.de>
> To: molpro-user at molpro.net
> Subject: [molpro-user] FCIQMC Error
> Message-ID: <1d0f2c05-f67c-bbb9-35f5-46d5327e5cb8 at fu-berlin.de>
> Content-Type: text/plain; charset=utf-8; format=flowed
> 
> Hello Molpro Users,
> 
> I want to test and run some FCIQMC calculations. To get started I took the first example from the manual (http://www.molpro.net/info/current/examples/nn_fciqmc.com) and submitted a molpro calculation
> (Version 2015.1). However I keep getting the following error message:
> 
> Fatal error in PMPI_Bcast: Invalid datatype, error stack:
> PMPI_Bcast(1610): MPI_Bcast(buf=0x3ca7c8c0, count=104, MPI_DATATYPE_NULL, root=0, MPI_COMM_WORLD) failed
> PMPI_Bcast(1570): Datatype for argument datatype is a null datatype
> 
> I also tested the other examples, without any success.
> This is very similar to a previously reported issue (http://www.molpro.net/pipermail/molpro-user/2017-July/006999.html): same error message but different numbers. However, this one did not receive any
> (public) answer.
> 
> Can anyone help me out?
> 
> best regards
> Christian Stemmle
> 
> ------------------------------
> _______________________________________________
> Molpro-user mailing list
> Molpro-user at molpro.net
> http://www.molpro.net/mailman/listinfo/molpro-user

--
Prof. Peter J. Knowles             
School of Chemistry, Cardiff University, Main Building, Park Place, Cardiff CF10 3AT, UK
Telephone: +44 29 208 79182
Email: KnowlesPJ at Cardiff.ac.uk 

Yr Athro Peter J. Knowles
Yr Ysgol Cemeg, Prifysgol Caerdydd,
Prif Adeilad, Plas y Parc
Caerdydd CF10 3AT, DU
Ffôn: +44 2920 879182
Ebost: KnowlesPJ at Caerdydd.ac.uk
Ysgrifennwch ataf yn Gymraeg neu Saesneg