[IPython-User] questions about IPython.parallel
Wed Oct 24 15:07:37 CDT 2012
thanks for the answer
2012/10/24 MinRK <firstname.lastname@example.org>:
> On Wed, Oct 24, 2012 at 3:36 AM, Francesco Montesano
> <email@example.com> wrote:
>> Dear list,
>> I have a bunch of coded designed to repeat the same operation over a
>> (possibly large)
>> number of file. So after discovering Ipython.parallel not long ago, I
>> decided to
>> rewrite to give me the possibility to use a task scheduler (I use
>> load_balance_view) in order
>> to make the best use possible of my quad core machines.
>> Here is the typical structure of my code
>> ###### BEGIN example.py ######
>> def command_line_parsing( ... ):
>> "in my case argparse"
>> def do_some_operation( ... ):
>> "executes some mathematical operation"
>> def read_operate_save_file( file, ... ):
>> """reads the file, does operations and save to an output file"""
>> input = np.loadtxt( file )
>>  do_some_operation( )
>> np.savetxt( outfile, ..... )
>> if __name__ == "__main__":
>> args = command_line_parsing( )
>> #parallelisation can be can chosen or not
>> if args.parallel :
>> #checks that Ipython is there, that an ipcluster has been started
>> #initialises a Client and a load_balance_view. I can pass a string
>> #list of strings to be executed on all engines (I use it to
>> "import xxx as x" )
>> lview = IPp.start_load_balanced_view( to_execute )
>> if( args.parallel == False ): #for serial computation
>>  for fn in args.ifname: #file name loop
>> output = read_operate_save_file(fn, dis, **vars(args) )
>> else: #I want parallel computation
>>  runs = [ lview.apply( read_operate_save_file,
>> os.path.abspath(fn.name), ... ) for fn in args.ifname ]
>> results = [r.result for r in runs]
>> ###### END example.py ######
>> I have two questions:
>>  In function 'read_operate_save_file', I call 'do_some_operation'. When
>> work on serial mode, everything works fine, but in parallel mode I get
>> the error
>> "IPython.parallel.error.RemoteError: NameError(global name
>> 'do_some_operation' is not defined)"
>> I'm not surprised by this, as I imagine that each engine know only what
>> has been
>> executed or defined before and that lview.apply( func, ... ) just passes
>> "func" to the engines. A solution that I see is to run "from example
>> do_some_operation" on the engines when initialising the load_balance_view.
>> there any easier/safer way?
> This namespace issue is common, and I have explanations scattered about the
> Which I really need to consolidate into a single thorough explanation with
> But the gist:
> - If a function is importable (e.g. in a module available both locally and
> remotely), then it's no problem
> - If it is defined in __main__ (e.g. in a script), then any references will
> be resolved in the *engine* namespace
> I recommend conforming to the first case if feasible, because then there
> should be no surprises.
> Everything surprising happens when you have depend on references in
> `__main__` or the current working dir (e.g. locally imported modules), since
> `__main__` is not the same on the various machines, nor is the working dir
> That said, if the names you need to resolve are few, a simple import/push
> step with a DirectView to set up namespaces should be all you need prior to
> submitting tasks (assuming new engines are not arriving in mid-computation).
> rc = Client()
> dv = rc[:]
> # push any locally defined functions that your task function uses:
> dv['do_some_operation'] = do_some_operation
I ended up doing the following when initialising the load_balance_view
dv.execute( 'import sys' )
dv.execute( 'sys.path.append("path_to_example.py")' )
dv.execute( 'from example import do_some_operation' )
Your suggestion looks much neater, just a couple of questions.
With the push that you suggest, do I simply call the
'do_some_operation' as in my example or do I need some different
Do you think that one or the other way is more optimal when the
function is called and executed?
> # perform any imports that are needed:
> dv.execute("import numpy as np...")
> # continue as before:
> lview = IPp.start_load_balanced_view( to_execute )
>>  Because of the way I parse my command line arguments, args.ifname its
>> list of already opened files. In serial mode, this is no problem, but when
>> assign the function to the scheduler passing the file, I get an error
>> that the cannot work on a closed file. If I pass the file name with the
>> absolute path, numpy can read it without problem. Is this a behaviour to
>> expected or a bug?
> I would expect a PickleError when you try to send an open file. Definitely
> send filenames, not open file objects.
Just a curiosity: what is the working directory of the engines? Is the
one where the ipcluster is started or where the profile is stored?
(While fixing my code, I ended up passing the filename with the full path)
>> Thanks for any help,
>> IPython-User mailing list
> IPython-User mailing list
More information about the IPython-User