Running MPI on 2 laptops
 

Greetings Ahsan,

Two reference posts that can come in handy:

• http://www.cfd-online.com/Forums/par...ple-cores.html
• http://www.cfd-online.com/Forums/ope...m-deux-pc.html

So, basically what happens in your case is two things:

1. The two laptops are not aware of the host names of each other, therefore are not able to find their corresponding IP address.
2. Your "machines" file is incorrect.

To solve the problem, you must first find the IP address for each laptop. You can use this command:
And then look for the IP address at "inet addr:" related to "eth0", "eth1" or "wlan0" or "wlan1".
Make a note of those IP addresses and try pinging the other machine, using it's IP address, to check if it can find it. For example:

• If "machine1" has the IP "11.12.13.14";
• and if "machine2" has the IP "11.12.13.132";

Then:

• From "machine2", run:
  
  Code:

  ---

  ping 11.12.13.14

  ---

• From "machine1", run:
  
  Code:

  ---

  ping 11.12.13.132

  ---

They should be able to see each other.

Now, once the IPs are properly determined and test, edit the file "/etc/hosts" as super-user in both machines and add the two lines at the end of the file:
Now, the other problem is the "machines" file you have got. It should be something like this:
Best regards,
Bruno

Thank you cat, i have successfully pinged both laptops, and did all the changes you mentioned.

Pinged:

My hosts file says:

But when i run:
:/.
Also when i ssh from ahsan: connection timed out
While when i ssh from umer : Permission denied

Hi Ahsan,

Sorry, I was not able to answer any sooner.
In summary, the problems in question are probably:

1. Do not use the @ symbol in the name of a machine. The one you are seeing in your command line is only implying that your user name is logged in at your machine; it does not mean that the whole text is the name of the machine.
   In your case, the actual name of your machine is only "ahsan-Inspiron-5521", so you should adjust the files accordingly.
2. The error message is indicating that no SSH connection was possible over the port 22. You need to configure SSH for it to work on both machines: https://help.ubuntu.com/community/SSH

Best regards,
Bruno

Help needed!
 

I have successfully SSH both laptops. But it gives error :

Cannot find executable s

Is it necessary for 2 laptops to have same version of ubuntu, Same openFOAM version, Same architecture (32 or 64 bit) to run a case across a network?
In our case one laptop is 32 and other 64 bit also 1 laptop has openfoam version 2.2.1 and other has 2.1.1. I think the problem lies in architecture and ubuntu version. Plz help me out on this.

Hi Ahsan,

Don't expect OpenFOAM to do universal parallel processing ;)
It doesn't matter what Ubuntu versions you have on each machine. But it does matter that you use the same exact architecture and version of OpenFOAM, as well as installed in the same path.

If one laptop is using a 32bit version of Ubuntu, then you have to install the same version of OpenFOAM and the same 32bit build, on both machines! This is because it's the common denominator between both machines, because 64bit machines can also use 32bit applications.

First choose which version of OpenFOAM you want to use and let me know, so that I can give you directions on what to do.

Best regards,
Bruno

cluster problem
 

Hey wyldcat i really appreciate your help. I have successfully run an airfoil3d case on two core i 3 machines. :) now i have moved to 3 machines. Actually what I did is that I installed same ubuntu(12.04 LTS) and openfoam (221) versions on all machines. Made same username and password account on all machines. All machines have same architecture. My host file looks like this:

ahsan cpu=4
alvi cpu=4
aftab cpu=4

In /etc/hosts I added these lines as you instructed:

192.168.1.2 ahsan
192.168.1.3 alvi
192.168.1.10 aftab

All machines ping and ssh successfully.
I start mpirun by typing:

mpirun -np 12 --hostfile ./hosts /opt/openfoam221/bin/foamExec simpleFoam -parallel

Now the problem is when i run my case on a single machine using its 4 cores it runs faster and reaches 200 time-steps in 5 mins 44 sec, while when I run it along 3 machines using 12 cores, it takes 10 min 33 sec. I used a hub to connect all pcs together. Decomposition method is "scotch". I decomposed my domain into 12 sub domains by typing "decomposePar" in all 3 machines. I dont know where the problem lies :(. Have i given enough info? plz tell me ill provide.

Hi Ahsan,

Well, throwing more CPUs at a problem doesn't mean that it will solve the problem faster. You need to take into account several details:

1. If all CPUs are equally powerful.
2. If the memory speed on all machines is the same.
3. If the connection between machines is powerful enough.

In addition, using Hyper-Threading isn't very good when using OpenFOAM. Therefore, if your machines are using i3 CPUs, then you should only use 2 cores on each one.


If you can describe:

1. The CPU model on each machine (specific model, e.g. i3-3115C);
2. The RAM speed on each machine, e.g. DDR3 1333 MHz;
3. The times that the ping command gives you, when checking the connection with the other machines;
4. The Ethernet connection, namely if it's 100 Mbps or 1 Gbps;
5. The number of cells in the mesh;

I can estimate what can be improved or if it will even work at all.

Best regards,
Bruno

Specs
 

intel (R) Core (TM) i3-2120 CPU @ 3.30GHz 3.30GHz
Installed Memory 8 Gb
64 bit operating system,x64 based processor
ethernet connection 100Mbps
RAM DDR3 1333MHz

No. of cells in mesh 1.1 million
mesh density (150 150 10)

check the attachment for ping times.

Hi Ahsan,

It has a Passmark index of 3871. Pretty powerful CPU! It's about 45% of an i7-2600. Which is pretty good, considering that it's 2 cores on the i3 vs 4 cores on the i7!

OK, no problem here.
Good.
Not good. Really not good :(
Any chance you can change to a Gigabit network, namely 1000Mbps = 1Gbps? Because this is the main reason why you're getting such bad timing results :(

For a sense of perspective: the CPUs would have to be running at 400 to 800 MHz per core, for you to notice improvements in using multiple machines. Which at this point, doesn't make much sense, because each CPU runs at 3300MHz.

Good enough!

150*150*10 = 225000
:confused: this only equates to 225 thousand cells.

0.700 ms is an indicator of pretty bad latency issues :( You need at least 0.300 to 0.100 to achieve relatively good simulation time.


If you cannot change the Ethernet connection to Gigabit, then you will have to resort to another kind of parallel performance: you can run 3 independent simulations at the same time, one of each machine, to test different settings.

If you can change the Ethernet connection to Gigabit and use 2 cores per machine, then I expect that you'll see something like this:

• 1 machine: 5 min
• 2 machines: 3 min 30 s
• 3 machines: 2 min 10 s

This is also depending on the number of cells per processor, which should be above 50000 to 100000 cells per processor.

Best regards,
Bruno

hi,ahsan~
should the case folder on different laptops be located in the same path ?
and where will the results be stored, the host or the slave?

thx~

Path
 

Case folder should have same path in all Pcs. If Pc1 is host then from processor 0-4 results will be stored in Pc1, 5-8 results will be stored in Pc2 and so on.:)

thank you for your vary fast reply!
another question~
is the password-less ssh necessary? i have not understood the detailed procedure even after reading this instruction on http://homepages.inf.ed.ac.uk/imurra...dless_ssh.html.
i am a freshman to openfoam as well as linux ~~

Besides, could you show me the " roots" part in decomposeParDict file?
i am not sure what the " node" means, one node means one PC ?

Ssh
 

I tried passwordless ssh but didnt work fr me. As far as my knowledge is concerned openfoam may not initiate MPI without a passwrd ssh. Il give u a key shortcut to remove all the fuss in making a cluster. Make same username account and password in all Pcs. Give different IPs to each Pc and run SSH. Pfff ! cluster is ready. Abt decomposePar ill ans that in a while.

help~
 

hi,bruno~
i have learnt a lot from your post, so really thank you for your work!
i have several questions about parallel running on two PCs(actually, two workstations).

1. you mentioned the username. So if two machines have the same username, is it still necessary for me to edit the "roots" in decomposeParDict ?
besides, as far as i am concerned, one node means one PC ? i have 2 workstations and each of them has 8 cores. so the " n" in " roots n" should be 1 or 15 ?

2. i still fail to understand how to get a password-less SSH even after reading the instructions on http://homepages.inf.ed.ac.uk/imurra...dless_ssh.html.
i am a freshman to linux~
is it possible to avoid the password-less SSH setting?

thank you again~

Greetings to all!

@ahliyu:
I suggest that you read the following pages:

• http://linuxcommand.org/learning_the_shell.php
• http://www.thegeekstuff.com/2008/11/...n-ssh-copy-id/
• Optional: Notes about running OpenFOAM in parallel

It's explained here: Running OpenFOAM in parallel with different locations for each process


Yes. Usually the nodes are referred to:

• master node - the main machine, responsible for keeping track of the jobs being executed in the cluster and usually also responsible for running part of the parallel case.
• slave nodes or worker nodes - as the name says, these nodes usually only do the heavy computational work, all together.

-------------------
edit: I moved the post above from the other quoted thread, to keep the discussion in the same thread here.

The "roots" are only a feature that allows that each processor folder is placed in different locations on each machine. If you do not define the "roots", the same exact path should be used for the main case folder and to somehow be shared among all machines.

Already answered above.

Also already answered above.

Best regards,
Bruno