Globals, classes and Parallel Python

Question

I have been reading for a few hours on how globals=globals() works on Parallel Python, and I am still a little bit confused, maybe you can help me... I am writing a code which can basically be summarized as:

import pp

class Foo(object):
    def __init__(self):
        self.h = 2

def f():
    foo = Foo()
    return foo.h

ppservers = ()
job_server = pp.Server(ppservers=ppservers)
#print globals()
g = job_server.submit(f, (), globals = globals())
r = g()

The output is "global name 'Foo' is not defined", as a few others have encountered before me... I know that the globals argument is used to transfer functions and classes to the server in a simplified way, so I hoped that it had passed the Foo() class, as it is in the global variables before executing the job_server.submit instruction (as you can see by de-commenting the print globals() instruction).

What am I missing about PP that I should know? Is there a way to make this code run without too many changes?

Thanks for reading me, I hope that some of you get how this works!

(Remark: I do not want to create an instance of Foo outside of the parallelized jobs, simply because initializing an instance triggers the whole program, which is within the init(self) function... bad idea, I know...)

score 0 · Accepted Answer · answered Feb 03 '15 at 18:06

This is an issue revolving around being able to have pp track down and associate the code dependencies correctly. pp inspects the first object passed in to submit (i.e. f), and extracts the source code… which it then passes to the other processes. It also passes in any additional objects that are passed in globals. However, pp primarily tracks functions, classes, and modules -- it has trouble with instances and many other objects.

If you use a fork of pp called ppft (it imports as pp), it should work as expected. ppft uses a better code inspection package (dill.source from dill), to extract the source code from a broader range of python objects and more complex dependencies.

>>> import pp
>>> class Foo(object):
...   def __init__(self):
...     self.h = 2
... 
>>> def f():
...   foo = Foo()
...   return foo.h
... 
>>> ppservers = ()
>>> job_server = pp.Server(ppservers=ppservers)
>>> g = job_server.submit(f, (), globals=globals())
>>> r = g()
>>> r
2

Get ppft here: https://github.com/uqfoundation

It's also pip installable.

Globals, classes and Parallel Python

1 Answers1