I found that some other lines of your output are doubled (for example some eigenvalues and headlines are printed twice). Now I think that the problem is as follows:

1. The "mpirun -np 8" runs 8 copies of g09 program.

2. Gaussian isn't an MPI program so all the eight copies do the same work (that would explain why there is no speedup in comparison to one-CPU job).

3. You use two nodes ("#PBS -lselect=2:ncpus=4") and they somehow compete when they write to the output file, so some parts of the file are doubled. I'm not sure how the mechanism works and why only certain lines are printed twice. You could rerun your calculation on three nodes with "#PBS -lselect=3:ncpus=4", "mpirun -np 12" and "%nprocshared=12" to see if "Normal termination" would be printed thrice.

The third point is just my guess, but with the first two I'm pretty sure. If you want to run your job in parallel, you are perhaps limited to one-node, many-cpu calculations ("#PBS -lselect=1:ncpus=4", "%nprocshared=4" and running g09 without "mpirun" command). There is something called Linda designed to run Gaussian on many nodes, but it isn't sold with g09 by default so you should ask your cluster administrator if it's installed. If you have access to it, it's controlled from your input file, without use of mpirun.

Calculations were performed in parallel. However, using 8 processors there is no speed increase when compared to the use of only one processor. I believe it has some error in the configuration of the program, but we could not identify what it is.

Thank you for your comment.

The first message appears on my output.

# # #

  % MEM = 1500MB

  nprocshared = 8

  Will use up to 8 processors via shared memory.

# # #

This other message does not appear.

Why is it?

The information "ntasks = XX, cpus / task = X, ncpus = XXX, nodes = XY" do not appear in my output.

You imagine what can be the problem?

Hi,

The calculations are performed on a cluster (supercomputer).

Although apparently the calculations are correct, it appears that the calculation isn´t being performed in parallel. There are times that the calculation using 8 processors takes longer than with just one. Sometimes, with 8 processors is just a bit faster than with only one processor.

PBS file is below. Suggests a change on it? Or should not interfere? The .PBS file was created by employees who manage the supercomputer.

Thanks.

====

#!/bin/bash

# ----------------------------MrBayes ------------------------------

# Configuration PBS

### Where XXXXXXX is put a name to your process

#PBS -N XXXXX

##### Set the number of us LSELECT = numbers of ncpus = numbers of cores

#PBS -lselect=2:ncpus=4

#PBS -m ae

#### Enter your email below

#PBS -M [email protected]

# ------------------------------Exemple --------------------------

# a - abort ; b - begin; e - exit

# -lrecurso=gauss -lrecurso=outrogauss -lgauss=4

# Adds date and time to the log file to determine the start time of execution

date

date

cd $PBS_O_WORKDIR

source /usr/share/modules/init/bash

export g09root=/home/progs

export GAUSS_SCRDIR=/tmp

source $g09root/g09/bsd/g09.profile

# Load the module

module load mpich2

# Set the hosts that are available

HOSTLIST=`cat $PBS_NODEFILE |paste -s -d ","`

cat $PBS_NODEFILE

# Put here your command

mpirun -np 8 /home/progs/g09/g09 m1.gjf

date

Are you sure that mpirun is the correct way to run G09 on multiple nodes? I perform only single node - many processor calculations with Gaussian, so I may be wrong. However, aren't all parallelization methods done by g09 itself (many-nodes support granted by Linda)? In that case execution of mpirun would cause all your nodes to perform the same job.

Stéphane..

Thanks for your attention.

All the best.

Fabricio

I'm just curious, is "Normal termination of Gaussian" the only part of your output that's doubled? Are there any redundant parts in the results, for example "TOTAL ENERGY" line printed twice? It's possible that two nodes do the same job and they even generate a latency when they attempt to access the same output file, which in turn increases the duration of the job.

It is very curious. Only the "Normal termination of Gaussian" is duplicated. You suggest something that might be happening? See anex.

Hi,

The calculation completed correctly. When I use only one processor, the duplicate message is not displayed.

I'll check if I have the job output file and attached here. If I do not have it, I rerun the calculation, since it is fast.

I appreciate the attention.

Fabricio.

Stéphane... What is this output file showing the number of CPUs used? Would not the output. .log?

I found that some other lines of your output are doubled (for example some eigenvalues and headlines are printed twice). Now I think that the problem is as follows:

1. The "mpirun -np 8" runs 8 copies of g09 program.

2. Gaussian isn't an MPI program so all the eight copies do the same work (that would explain why there is no speedup in comparison to one-CPU job).

3. You use two nodes ("#PBS -lselect=2:ncpus=4") and they somehow compete when they write to the output file, so some parts of the file are doubled. I'm not sure how the mechanism works and why only certain lines are printed twice. You could rerun your calculation on three nodes with "#PBS -lselect=3:ncpus=4", "mpirun -np 12" and "%nprocshared=12" to see if "Normal termination" would be printed thrice.

The third point is just my guess, but with the first two I'm pretty sure. If you want to run your job in parallel, you are perhaps limited to one-node, many-cpu calculations ("#PBS -lselect=1:ncpus=4", "%nprocshared=4" and running g09 without "mpirun" command). There is something called Linda designed to run Gaussian on many nodes, but it isn't sold with g09 by default so you should ask your cluster administrator if it's installed. If you have access to it, it's controlled from your input file, without use of mpirun.

Pavel, thank you!

You solved my problem.

We did a test like this:

So we changed the GJF file to:

#nprocshared=8

and PBS file for:

-lselect=1: ncpus=8

And even while maintaining the below command worked properly and faster.

mpirun-np 1 / home/progs/g09/g09 m1.gjf

Thank you. Let's do some more tests still best to check the amount of memory, for example.

By their experience, number of processors equal to 8 is a good number?

I'm glad I could help.

I think that when you use just one node, you can get rid of the "mpirun" command entirely. It uses one computer anyway. As for the number of processors, I'm limited to four, but just because most nodes on our server have that many. The best solution would be to run a job several times with increasing number of cpus to find what the speedup is, as it's highly program and hardware dependent. I've heard that Gaussian doesn't scale well above eight processors, but I've never run such tests.