parallel-examples

Examples of parallel code using ipython on ascar

setting up ascar (one time stuff)

• Edit your ~/.bashrc or ~/.bash_profile to include the correct version of python

  We have two versions of python installed on ascar. Enthought Canopy and Continuum's Anaconda. Currently Enthought Canopy has a working version of the GIS python modules, Anaconda does not.

  ONLY USE ONE OF THE BELOW

  To activate Enthought Canopy use the following source /share/apps/enthought/User/bin/activate To activate Continuum Anaconda use the following export PATH="/share/apps/anaconda-1.6/bin:$PATH

  Currently I have only tested Canopy with the parallel code

• setup your basic parallel ipython environment

  • $ ipython profile create sge --parallel

  • Edit ~/.config/ipython/profile_sge/ipcontroller_config.py, adding the line:

    <code> c.HubFactory.ip = '*' </code>
    
    to instruct the controller to listen on all interfaces.
    

  • Edit ~/.config/ipython/profile_sge/ipcluster_config.py, adding the lines:

    <code> c.IPClusterEngines.engine_launcher_class = 'SGEEngineSetLauncher'</code>
    
    <code> c.IPClusterStart.controller_launcher_class = 'SGEControllerLauncher'</code>
    
    <code> c.SGELauncher.queue = 'all.q'</code>
    

    at this point you should be able to start a cluster using

    $ ipcluster start -n 10 --profile=sge --cluster-id=test

    check that is started using

    $ qstat

    and stop it using

    $ ipcluster stop --profile=sge --cluster-id=test

installing ascar-parallel

This package ascar-parallel does some of the boiler-plate of starting and stopping the cluster for you.

It is already installed on the canopy python but for reference, to install:

$ git clone [email protected]:twdb/ascar-parallel.git

$ cd ascar-parallel

$ python setup.py install

using ascar parallel

in your code use the following:

from ascar_parallel import StartCluster

with StartCluster(8) as lview:
	lview.map(myfunc, <args>)

more details in the examples folder

common issues

• imports The simplest way to make imported packages available to all the compute engines you can import them inside the function you are parallelizing, i.e.

   from ascar_parallel import StartCluster
  
   def myfunc(a,b):
   	import numpy as np
   	return np.sqrt(a+b)
  
   a_list = range(5)
   b_list = range(5)
   with StartCluster(8) as lview:
   	lview.map(myfunc, a_list, b_list)

  if you are defining a bunch of helper functions for myfunc, then these will not be available on the compute engines. You can do one of two things.

  • define the functions inside myfunc:

   def myfunc(a,b)
   	def helperfunc(c,d):
   		stuff
  
   	<statements>

  • put the helper functions inside another file/module and then import them. This is the prefered way

   def myfunc(a,b)
   	from mymodule import helperfunc
   	<statements>

• files: ascar has four main shared folders:

  • /home - user home directories, this is backed up. DO NOT SAVE MODEL RESULTS HERE
  • /share/swr - this is backed up. About 500GB of space currently available
  • /share/work - this is not backed up. About 7TB of space currently available
  • /share/apps - this is a good place to put software and executable code (i.e. fortran etc). You can then add this to your path so that executables are available from all the compute engines.