Re: [ESPResSo-devel] [bug #37281] cannot read blockfile with mpirun

[Stefan Kesselheim]

Dear Martin,
thanks for submitting this problem.
It is not a bug of Espresso (see below) and in these cases, or if you are unsure, please send a mail to the espresso users mailing list: address@hidden.
This list is read by more people, and often other users can help out.
For your problem: You are using the variable name "in" as input file channel. In the blockfile, there is a variable "in" defined. When reading this variable, you destroy the input channel.
It's a bit tricky, but you will have to make sure that there are no variable name collisions.
I'm no sure how you found the diagnosis that the problem being related to MPI, because in principle MPI and blockfiles are fully orthogonal: The blockfile contains information interpreted by the TCL interpreter and this is only done in the process with rank 0, and all its content is then interpreted on the TCL level serially.

Finally: A smaller blockfile (that is not compressed) would have saved time in finding the reason :-). I merely stumbled over it by a similar name collision trying to reproduce you problem.

Good luck and cheers
Stefan

On 09/06/2012 11:18 AM, Martin Linden wrote:

URL:
  <http://savannah.nongnu.org/bugs/?37281>

                 Summary: cannot read blockfile with mpirun
                 Project: ESPResSo
            Submitted by: bmelinden_dbb
            Submitted on: Thu 06 Sep 2012 09:18:34 AM GMT
                Category: Simulation core
                Severity: 3 - Normal
                  Status: None
             Assigned to: None
             Open/Closed: Open
         Discussion Lock: Any
                 Release: 3.1.0
           Fixed Release: None

    _______________________________________________________

Details:

I have trouble reading in blockfiles when using mpi. 
I would say this is a pretty serious problem, because it means that I can only
restart crashed simulations on a single processor.

The attached script and gzip archive demonstrates the problem. Without MPI,
Espresso reads the block file and starts integrating:

$ Espresso blockread.tcl

With mpi, the fine is somehow not opened, or not recognized:

$ mpirun -n 4 Espresso blockread.tcl

(...)

can not find channel named "file14"
    while executing
"blockfile $in read auto"
    invoked from within
"while { [blockfile $in read auto] != "eof" } {}"
    (file "blockread.tcl" line 6)
--------------------------------------------------------------------------
mpirun noticed that the job aborted, but has no info as to the process
that caused that situation.
--------------------------------------------------------------------------

The block file in question was in fact created using mpi, so writing blocks
seems to work.

System and version:
ubuntu 12-04 64 bit, espresso 3.1.0, code_info: { Compilation status { FFTW }
{ BOND_ANGLE_HARMONIC } { LENNARD_JONES } { LJCOS } { LJCOS2 } { EXCLUSIONS }
}

Best,

Martin Lindén, Stockholm University




    _______________________________________________________

File Attachments:


-------------------------------------------------------
Date: Thu 06 Sep 2012 09:18:34 AM GMT  Name: blockread.tcl  Size: 1kB   By:
bmelinden_dbb

<http://savannah.nongnu.org/bugs/download.php?file_id=26486>
-------------------------------------------------------
Date: Thu 06 Sep 2012 09:18:34 AM GMT  Name: checkpoint.gz  Size: 170kB   By:
bmelinden_dbb

<http://savannah.nongnu.org/bugs/download.php?file_id=26487>

    _______________________________________________________

Reply to this item at:

  <http://savannah.nongnu.org/bugs/?37281>

_______________________________________________
  Message sent via/by Savannah
  http://savannah.nongnu.org/