Hello MPI in Python
This tutorial will demonstrate how to use Bryn Mawr College's Athena Cluster with Python and MPI, the Message Passing Interface.
Here are the steps, briefly, for running a MPI Python Script on Athena
- log into athena.brynmawr.edu
- find out which /data path you have a shared directory (either /data1, or /data2)
- create a python program in your /dataN/userid/ directory
- create a PBS script
- submit the PBS script
- check on status
- see the results
We will now go through each step in detail.
Step 1: Log into Athena
If you do not have an Athena account, please email ... to request one. It can take up to one business day.
After you have your account, you can log into the account from anywhere. You can use ssh on Linux and Mac, or putty on Windows.
You can use the X-forwarding flag, -X, if you have X running:
ssh -X firstname.lastname@example.org
Using X-forwarding will allow you to use graphical programs, such as gedit.
Step 2: Where is my /data?
Next, you need to find out where your /data directory (either /data1 or /data2) has been located. The easiest way is to list the files in the directories:
[dblank@powerwulf ~]$ ls /data1 dblank gdavis n001.cluster.com n004.cluster.com pbrodfue dmtpc josh n002.cluster.com n005.cluster.com powerwulf.brynmawr.edu emaccorma mrice n003.cluster.com n006.cluster.com xwang [dblank@powerwulf ~]$ ls /data2 dblank gdavis n001.cluster.com n004.cluster.com pbrodfue emaccorma josh n002.cluster.com n005.cluster.com powerwulf.brynmawr.edu galaxy mrice n003.cluster.com n006.cluster.com xwang [dblank@powerwulf ~]$
Step 3: Create a Python script
Next, you will want to create a Python script in your /data directory.
First, change to your directory:
Next, use a text editor (emacs, vi, gedit, or nano) to create a file in your /data directory. We use the /data directory because it is available to all of the nodes.
Here is a sample program:
# More tutorials available at: # http://mpi4py.scipy.org/docs/usrman/tutorial.html from mpi4py import MPI import platform comm = MPI.COMM_WORLD rank = comm.Get_rank() size = comm.Get_size() machine = platform.node() print("Hello MPI from %s %d of %d" % (machine, rank, size))
Save the script to a file, such as: /data1/dblank/hellompi.py
rank and size are two useful values from MPI. See http://mpi4py.scipy.org/docs/usrman/tutorial.html for more examples.
Step 4: Submit PBS script
To submit a Python script, you will need a shell program to give to qsub. Here is a sample:
#!/bin/bash #PBS -q default #PBS -l nodes=6:ppn=12 mpiexec -machinefile /opt/machinelist -n 10 python /data1/dblank/hellompi.py echo "end"
Save this in a file, such as /home/dblank/pbs/hellompi.sh.
PBS is a language that allows you to set many aspects of your running program. For more details see ...
In this example, we tell mpiexec to choose from the machines listed in /opt/machinelist (all of the nodes) and to create 10 instances of our running code across those machines.
NOTE: the program will start in your home directory, so by default you will be in /home/dblank/. Use full pathnames to /data1/ for shared data.
Finally you will submit this script to qsum, like so:
NOTE: the results of qsub will appear here. Therefore, you may want to move your shell script to a folder before beginning.
Step 5: Check status
When you start your program a number of things will happen:
- your program will be scheduled to run
- when qsub runs your script, the script be submitted to each of the machines, perhaps a number of times
You can run "qstat -a" to see the status of your program.
Step 6: Get Results
Finally, any output from your program will be saved in a file with the same name as your shell script, with a dot, the letter "o", and job number appended.
Hello MPI from n002.cluster.com 1 of 10 Hello MPI from n002.cluster.com 7 of 10 Hello MPI from n001.cluster.com 6 of 10 Hello MPI from n001.cluster.com 0 of 10 Hello MPI from n004.cluster.com 3 of 10 Hello MPI from n004.cluster.com 9 of 10 Hello MPI from n003.cluster.com 2 of 10 Hello MPI from n003.cluster.com 8 of 10 Hello MPI from n006.cluster.com 5 of 10 Hello MPI from n005.cluster.com 4 of 10 end