Author Topic: [SOLVED] Run numforce in SMP machine  (Read 4857 times)

Glxblt76

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
[SOLVED] Run numforce in SMP machine
« on: March 25, 2015, 08:38:12 am »
 Hi,

I have a 1-cpu machine with 8 cores (SMP machine). I just need Numforce to check if my lowest vibrational frequency is not imaginary. Thus, I do not need to calculate all numerical frequencies from Numforce, but only the lowest one.

I scripted Numforce in order to launch jobs one after the other, but I do not know how:

- To tell Numforce to use more than one node from the command;
- To tell Numforce to compute the lowest eigenvalue only.

The command I use in my script for the moment is simply:

nohup Numforce > Numforce.out &

What do I have to add in that command to perform my calculations on more than one core and to compute the lowest eigenvalue only? I red some of this forum's posts, so I guess that part of the answer lies in the -mfile option, but I was unable to find a complete account about how to use it in my Turbomole manual, so I'm a bit puzzled.

Best regards
« Last Edit: April 20, 2015, 02:03:34 pm by Glxblt76 »

antti_karttunen

  • Sr. Member
  • ****
  • Posts: 201
  • Karma: +1/-0
Re: Run numforce in SMP machine
« Reply #1 on: March 25, 2015, 12:36:54 pm »
Hi,

I think that the lowest eigenvalue search is only available in aoforce. With NumForce, you will always have to calculate the full hessian (except for possibly freezing atoms with -frznuclei)

For running the parallel jobs, you only need to create the machinefile and give it as a parameter. For your configuration, this is simple, just list the name of your machine eight times:

mymachine
mymachine
mymachine
mymachine
mymachine
mymachine
mymachine
mymachine


NumForce will make a remote connection to the machine via ssh, so you need to be able to log in without a password (e.g. by setting up a key-based authentication).

In addition to -mfile, it's important to set the -scrpath parameter so that NumForce will run in the correct temporary directory. Here you can use /tmp or whatever you wish. After all this, you can run NumForce with:
nohup Numforce -mfile my_m_file -scrpath /tmp > Numforce.out &

Best,
Antti

Glxblt76

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
Re: Run numforce in SMP machine
« Reply #2 on: March 25, 2015, 01:49:34 pm »
Thanks for your answer. I will test it when I will have some time and calculations to do it.

The answer may seem obvious to you, but can we assume that Numforce can be trivially parallelized? Since it does one single point for every frequency calculation, if I well understood?

Regards

antti_karttunen

  • Sr. Member
  • ****
  • Posts: 201
  • Karma: +1/-0
Re: Run numforce in SMP machine
« Reply #3 on: March 25, 2015, 01:56:06 pm »
Hi,

You are exactly right, NumForce is "embarrassingly parallel"!

Best,
Antti

Glxblt76

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
Re: [SOLVED] Run numforce in SMP machine
« Reply #4 on: April 20, 2015, 11:40:32 am »
Hi,

I tested and I have an issue.

I have a mfile, and a simple script looping on all subdirectories of my directory.

If I use the following line:

Code: [Select]
nohup NumForce -mfile ../mfile -scrpath /tmp > NumForce.out &
for each subdirectory, I get this error in my NumForce.out file:

Code: [Select]
ssh: connect to host theophile-HP-Z420-Workstation port 22: Connection refused^M

Error logging in on theophile-HP-Z420-Workstation:1

Thus, I expected that by using sudo command, I could remove that annoying permission problem:

Code: [Select]
sudo nohup NumForce -mfile ../mfile -scrpath /tmp > NumForce.out &
But all I get doing that is an empty NumForce.out file and the following error appearing directly in the command prompt:

Code: [Select]
nohup: failed to run command ‘NumForce’: No such file or directory
In short, I can't run my frequency calculations because of permissions problems, but when I allow myself the permissions, I have a weird bug preventing me from running calculations.

What to do? I tried to check if the NumForce script is in Windows format (I saved the default file, copied it, converted the copied file using dos2unix and tried to use it), but it didn't work. I suspect that there is some format problem somewhere (not in NumForce), but I don't know where it is and how to find it.
« Last Edit: April 20, 2015, 11:43:36 am by Glxblt76 »

Glxblt76

  • Jr. Member
  • **
  • Posts: 20
  • Karma: +0/-0
Re: [UNSOLVED] Run numforce in SMP machine
« Reply #5 on: April 20, 2015, 02:03:23 pm »
Okay I double-post because it seems I solved my own problem!

If anyone has this problem, that is how I proceeded:

I first noticed that I did not have ssh installed on my machine. So:
Code: [Select]
sudo apt-get install ssh
Then, I ran into a new problem: ssh asked me to enter my password during the execution of my script over all subdirectories, which was not possible. After struggling to understand how to automatically enter the password so I wouldn't be bothered and I could continue my work, I found >> this link << about Public-Key Based Authentication which really saved my life.

This was actually pointed out by antti_karttunen in a previous post on this topic but I wasn't aware that I didn't solved that problem before, since I thought it was already implemented in my machine.

After suitably configurating the few parameters for a Public-Key Based Authentication as indicated in the above link, I launched again my script by simply using the original expression without sudo:

Code: [Select]
nohup NumForce -mfile ../mfile -scrpath /tmp > NumForce.out &
and it worked nicely!  :D
« Last Edit: April 20, 2015, 02:22:31 pm by Glxblt76 »

himansu

  • Newbie
  • *
  • Posts: 4
  • Karma: +0/-0
Run CCSDT in SMP machine
« Reply #6 on: November 13, 2015, 04:10:13 am »
Dear Users,
    I want to do CCSDT calculation with SMP. I am using Turbomole 6.5, installed in a server, we do not use job submission script, instead we directly submit the job. I could manage to do Numforce and jobex calculation by using following commands

nohup Numforce -mfile ./hosts force.out &
nohup jobex -np 8  > jobex.out &

But unable to run ccsd(t) by using

nohup ricc2  -mfile ./hosts ricc2.out &
or
nohup ricc2 -np 8  > ricc2.out &

Waiting for your suggestions in this regard. By using job submission script in our hpc cluster I am able to SMP CCSDT calculation. Our job submission script looks like

export MPI_IC_ORDER="TCP"
export TURBODIR=/opt/intel/apps/TURBOMOLE-6-5/TURBOMOLE
source $TURBODIR/Config_turbo_env
export PARA_ARCH=SMP
export PARNODES=16
export PATH=$TURBODIR/scripts:$PATH
export PATH=$TURBODIR/bin/`sysname`:$PATH

Suggestions regarding SMP CCSDT calculation in a server are requested.

Regards,
Himansu

christof.haettig

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 258
  • Karma: +0/-0
    • Hattig's Group at the RUB
Re: [SOLVED] Run numforce in SMP machine
« Reply #7 on: November 20, 2015, 03:58:31 pm »
First of all TM 6.5 is outdated. You should switch to the new version...

If you want to run SMP parallel you need to have these settings:
export PARA_ARCH=SMP
export PATH=$TURBODIR/scripts:$PATH
export PATH=$TURBODIR/bin/`sysname`:$PATH

You can type them in on your shell everytime you want to start a calculation. Or put them in the profile of you shell (probably
~/.profile or ~/.bashrc, but if you don't know ask you sysadmin).

Christof