Thunderhead Engineering Forum

Please login or register.

Login with username, password and session length
Advanced search  

News:

Forum moved to https://forum.thunderheadeng.com

Author Topic: Parallel Processing  (Read 2937 times)

pkeogh

  • New Member
  • *
  • Posts: 2
    • View Profile
Parallel Processing
« on: October 03, 2012, 07:29:59 am »

Hello, I have created a model with 2 meshes and I was wondering if it is possible to run the model on multiple processors using I.P. addresses.  Previously, I was able to run a model on two processors using our network domain but our system has changed slightly and I can't seem to find any other way to link the two processors.

My first question, "is it possible to run multiple processing using I.P. addresses?" and secondly, "is there a specific way to input the I.P. addresses into the 'config.txt' file, and is there any other requirements for this type of multiple processing?".

I have tried numerous different ways to get this running but I keep getting the following error:
Unable to connect to 'IP Address',
sock error: Error = -1

Any help would be greatly appreciated.
Logged

Charlie Thornton

  • Thunderhead
  • *****
  • Posts: 851
    • View Profile
Re: Parallel Processing
« Reply #1 on: October 03, 2012, 09:20:01 am »

I can think of no reason why you wouldn't be able to enter an IP address instead of a host name, but I have been unable to verify that after a little bit of web searching. It looks like MPI knows you entered an IP address and that is encouraging. If you think the IP addresses are a problem, you can set up your computer to resolve hostnames manually. On Windows, you edit the lmhosts file and on linux you edit /etc/hosts.

If I were debugging the "unable to connect" message, I would:

1. Test running a simple problem on the local machine only, using "localhost"
2. Test running a simple problem on the local machine only, using its IP address
3. Repeat tests 1 and 2 on one of the other machines
4. Verify that both/all cluster nodes are using the same version of MPICH2
5. Verify that all firewalls know to let MPI through on the ports you've set it up to use

In making PyroSim's cluster feature we occasionally needed to drop back to the console and run MPI manually - to test things like computers on the same domain vs not, different account permissions, etc. It was always frustrating and seemed to break the moment our back was turned. My advice is to build up to a working configuration very scientifically.
Logged