I have a Python module that uses multiprocessing. I'm executing this module from another script with runpy. However, this results in (1) the module running twice, and (2) the multiprocessing jobs never finish (the script just hangs).
In my minimal working example, I have a script runpy_test.py:
import runpy
runpy.run_module('module_test')
and a directory module_test containing an empty __init__.py and a __main__.py:
from multiprocessing import Pool
print 'start'
def f(x):
return x*x
pool = Pool()
result = pool.map(f, [1,2,3])
print 'done'
When I run runpy_test.py, I get:
start
start
and the script hangs.
If I remove the pool.map call (or if I run __main__.py directly, including the pool.map call), I get:
start
done
I'm running this on Scientific Linux 7.6 in Python 2.7.5.
Rewrite your __main__.py like so:
from multiprocessing import Pool
from .implementation import f
print 'start'
pool = Pool()
result = pool.map(f, [1,2,3])
print 'done'
And then write an implementation.py (you can call this whatever you want) in which your function is defined:
def f(x):
return x*x
Otherwise you will have the same problem with most interfaces in multiprocessing, and independently of using runpy. As #Weeble explained, when Pool.map tries to load the function f in each sub-process it will import <your_package>.__main__ where your function is defined, but since you have executable code at module-level in __main__ it will be re-executed by the sub-process.
Aside from this technical reason, this is also better design in terms of separation of concerns and testing. Now you can easily import and call (including for test purposes) the function f without running it in parallel.
Try defining your function f in a separate module. It needs to be serialised to be passed to the pool processes, and then those processes need to recreate it, by importing the module it occurs in. However, the __main__.py file it occurs in isn't a module, or at least, not a well-behaved one. Attempting to import it would result in the creation of another Pool and another invocation of map, which seems like a recipe for disaster.
Although not the "right" way to do it, one solution that ended up working for me was to use runpy's _run_module_as_main instead of run_module. This was ideal for me since I was working with someone else's code and required the fewest changes.
Related
I'm trying to parallelize a script, but for an unknown reason the kernel just freeze without any errors thrown.
minimal working example:
from multiprocessing import Pool
def f(x):
return x*x
p = Pool(6)
print(p.map(f, range(10)))
Interestingly, all works fine if I define my function in another file then import it. How can I make it work without the need of another file?
I work with spyder (anaconda) and I have the same result if I run my code from the windows command line.
This happens because you didn't protect your "procedural" part of the code from re-execution when your child processes are importing f.
They need to import f, because Windows doesn't support forking as start method for new processes (only spawn). A new Python process has to be started from scratch, f imported and this import will also trigger another Pool to be created in all child-processes ... and their child-processes and their child-processes...
To prevent this recursion, you have to insert an if __name__ == '__main__': -line between the upper part, which should run on imports and a lower part, which should only run when your script is executed as the main script (only the case for the parent).
from multiprocessing import Pool
def f(x):
return x*x
if __name__ == '__main__': # protect your program's entry point
p = Pool(6)
print(p.map(f, range(10)))
Separating your code like that is mandatory for multiprocessing on Windows and Unix-y systems when used with 'spawn' or 'forkserver' start-method instead of default 'fork'. In general, start-methods can be modified with multiprocessing.set_start_method(method).
Since Python 3.8, macOS also uses 'spawn' instead of 'fork' as default.
It's general a good practice to separate any script in upper "definition" and lower "execution as main", to make code importable without unnecessary executions of parts only relevant when run as top level script. Last but not least, it facilitates understanding the control flow of your program when you don't intermix definitions and executions.
I have not been able to implement the suggestion here: Applying two functions to two lists simultaneously.
I guess it is because the module is imported by another module and thus my Windows spawns multiple python processes?
My question is: how can I use the code below without the if if __name__ == "__main__":
args_m = [(mortality_men, my_agents, graveyard, families, firms, year, agent) for agent in males]
args_f = [(mortality_women, fertility, year, families, my_agents, graveyard, firms, agent) for agent in females]
with mp.Pool(processes=(mp.cpu_count() - 1)) as p:
p.map_async(process_males, args_m)
p.map_async(process_females, args_f)
Both process_males and process_females are fuctions.
args_m, args_f are iterators
Also, I don't need to return anything. Agents are class instances that need updating.
The reason you need to guard multiprocessing code in a if __name__ == "__main__" is that you don't want it to run again in the child process. That can happen on Windows, where the interpreter needs to reload all of its state since there's no fork system call that will copy the parent process's address space. But you only need to use it where code is supposed to be running at the top level since you're in the main script. It's not the only way to guard your code.
In your specific case, I think you should put the multiprocessing code in a function. That won't run in the child process, as long as nothing else calls the function when it should not. Your main module can import the module, then call the function (from within an if __name__ == "__main__" block, probably).
It should be something like this:
some_module.py:
def process_males(x):
...
def process_females(x):
...
args_m = [...] # these could be defined inside the function below if that makes more sense
args_f = [...]
def do_stuff():
with mp.Pool(processes=(mp.cpu_count() - 1)) as p:
p.map_async(process_males, args_m)
p.map_async(process_females, args_f)
main.py:
import some_module
if __name__ == "__main__":
some_module.do_stuff()
In your real code you might want to pass some arguments or get a return value from do_stuff (which should also be given a more descriptive name than the generic one I've used in this example).
The idea of if __name__ == '__main__': is to avoid infinite process spawning.
When pickling a function defined in your main script, python has to figure out what part of your main script is the function code. It will basically re run your script. If your code creating the Pool is in the same script and not protected by the "if main", then by trying to import the function, you will try to launch another Pool that will try to launch another Pool....
Thus you should separate the function definitions from the actual main script:
from multiprocessing import Pool
# define test functions outside main
# so it can be imported withou launching
# new Pool
def test_func():
pass
if __name__ == '__main__':
with Pool(4) as p:
r = p.apply_async(test_func)
... do stuff
result = r.get()
Cannot yet comment on the question, but a workaround I have used that some have mentioned is just to define the process_males etc. functions in a module that is different to where the processes are spawned. Then import the module containing the multiprocessing spawns.
I solved it by calling the modules' multiprocessing function within "if __ name__ == "__ main__":" of the main script, as the function that involves multiprocessing is the last step in my module, others could try if aplicable.
I'm trying to run a few independent computations (though reading from the same data). My code works when I run it on Ubuntu, but not on Windows (windows server 2012 R2), where I get the error:
'module' object has no attribute ...
when I try to use multiprocessing.Pool (it appears in the kernel console, not as output in the notebook itself)
(And I've already made the mistake of defining the function AFTER creating the pool, and I've also corrected it, that's not the problem).
This happens even on the simplest of examples:
from multiprocessing import Pool
def f(x):
return x**2
pool = Pool(4)
for res in pool.map(f,range(20)):
print res
I know that it needs to be able to import the module (and I have no idea how this works when working in the notebook), and I've heard of IPython.Parallel, but I have been unable to find any documentation or examples.
Any solutions/alternatives would be most welcome.
I would post this as a comment since I don't have a full answer, but I'll amend as I figure out what is going on.
from multiprocessing import Pool
def f(x):
return x**2
if __name__ == '__main__':
pool = Pool(4)
for res in pool.map(f,range(20)):
print(res)
This works. I believe the answer to this question is here. In short, the subprocesses do not know they are subprocesses and are attempting to run the main script recursively.
This is the error I am given, which gives us the same solution:
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.
This probably means that you are not using fork to start your
child processes and you have forgotten to use the proper idiom
in the main module:
if __name__ == '__main__':
freeze_support()
...
The "freeze_support()" line can be omitted if the program
is not going to be frozen to produce an executable.
I have a program that, among other things, parses some big files, and I would like to have this done in parallel to save time.
The code flow looks something like this:
if __name__ == '__main__':
obj = program_object()
obj.do_so_some_stuff(argv)
obj.field1 = parse_file_one(f1)
obj.field2 = parse_file_two(f2)
obj.do_some_more_stuff()
I tried running the file parsing methods in separate processes like this:
p_1 = multiprocessing.Process(target=parse_file_one, args=(f1))
p_2 = multiprocessing.Process(target=parse_file_two, args=(f2))
p_1.start()
p_2.start()
p_1.join()
p_2.join()
There are 2 problems here. One is how to have the separate process modify the filed, but more importantly, forking the process duplicates my whole main! I get exception regarding argv when executing the
do_so_some_stuff(argv)
second time. That really is not what I wanted. It even happened when I run only 1 of the Processes.
How could I get just the file parsing methods to run in parallel to each other, and then continue back with main process like before?
Try putting the parsing methods in a separate module.
First, i guess instead of:
obj = program_object()
program_object.do_so_some_stuff(argv)
you mean:
obj = program_object()
obj.do_so_some_stuff(argv)
Second, try using threading like this:
#!/usr/bin/python
import thread
if __name__ == '__main__':
try:
thread.start_new_thread( parse_file_one, (f1) )
thread.start_new_thread( parse_file_two, (f2) )
except:
print "Error: unable to start thread"
But, as pointed out by Wooble, depending on the implementation of your parsing functions, this might not be a solution that executes truly in parallel, because of the GIL.
In that case, you should check the Python multiprocessing module that will do true concurrent execution:
multiprocessing is a package that supports spawning processes using an
API similar to the threading module. The multiprocessing package
offers both local and remote concurrency, effectively side-stepping
the Global Interpreter Lock by using subprocesses instead of threads.
Due to this, the multiprocessing module allows the programmer to fully
leverage multiple processors on a given machine.
When using multiprocessing.Pool in python with the following code, there is some bizarre behavior.
from multiprocessing import Pool
p = Pool(3)
def f(x): return x
threads = [p.apply_async(f, [i]) for i in range(20)]
for t in threads:
try: print(t.get(timeout=1))
except Exception: pass
I get the following error three times (one for each thread in the pool), and it prints "3" through "19":
AttributeError: 'module' object has no attribute 'f'
The first three apply_async calls never return.
Meanwhile, if I try:
from multiprocessing import Pool
p = Pool(3)
def f(x): print(x)
p.map(f, range(20))
I get the AttributeError 3 times, the shell prints "6" through "19", and then hangs and cannot be killed by [Ctrl] + [C]
The multiprocessing docs have the following to say:
Functionality within this package requires that the main module be
importable by the children.
What does this mean?
To clarify, I'm running code in the terminal to test functionality, but ultimately I want to be able to put this into modules of a web server. How do you properly use multiprocessing.Pool in the python terminal and in code modules?
Caveat: Multiprocessing is the wrong tool to use in the context of web servers like Django and Flask. Instead, you should use a task framework like Celery or an infrastructure solution like Elastic Beanstalk Worker Environments. Using multiprocessing to spawn threads or processes is bad because it gives you no oversight or management of those threads/processes, and so you have to build your own failure detection logic, retry logic, etc. At that point, you are better served by using an off-the-shelf tool that is actually designed to handle asynchronous tasks, because it will give you these out of the box.
Understanding the docs
Functionality within this package requires that the main module be importable by the children.
What this means is that pools must be initialized after the definitions of functions to be run on them. Using pools within if __name__ == "__main__": blocks works if you are writing a standalone script, but this isn't possible in either larger code bases or server code (such as a Django or Flask project). So, if you're trying to use Pools in one of these, make sure to follow these guidelines, which are explained in the sections below:
Initialize Pools inside functions whenever possible. If you have to initialize them in the global scope, do so at the bottom of the module.
Do not call the methods of a Pool in the global scope.
Alternatively, if you only need better parallelism on I/O (like database accesses or network calls), you can save yourself all this headache and use pools of threads instead of pools of processes. This involves the completely undocumented:
from multiprocessing.pool import ThreadPool
It's interface is exactly the same as that of Pool, but since it uses threads and not processes, it comes with none of the caveats that using process pools do, with the only downside being you don't get true parallelism of code execution, just parallelism in blocking I/O.
Pools must be initialized after the definitions of functions to be run on them
The inscrutable text from the python docs means that at the time the pool is defined, the surrounding module is imported by the threads in the pool. In the case of the python terminal, this means all and only code you have run so far.
So, any functions you want to use in the pool must be defined before the pool is initialized. This is true both of code in a module and code in the terminal. The following modifications of the code in the question will work fine:
from multiprocessing import Pool
def f(x): return x # FIRST
p = Pool(3) # SECOND
threads = [p.apply_async(f, [i]) for i in range(20)]
for t in threads:
try: print(t.get(timeout=1))
except Exception: pass
Or
from multiprocessing import Pool
def f(x): print(x) # FIRST
p = Pool(3) # SECOND
p.map(f, range(20))
By fine, I mean fine on Unix. Windows has it's own problems, that I'm not going into here.
Using pools in modules
But wait, there's more (to using pools in modules that you want to import elsewhere)!
If you define a pool inside a function, you have no problems. But if you are using a Pool object as a global variable in a module, it must be defined at the bottom of the page, not the top. Though this goes against most good code style, it is necessary for functionality. The way to use a pool declared at the top of a page is to only use it with functions imported from other modules, like so:
from multiprocessing import Pool
from other_module import f
p = Pool(3)
p.map(f, range(20))
Importing a pre-configured pool from another module is pretty horrific, as the import must come after whatever you want to run on it, like so:
### module.py ###
from multiprocessing import Pool
POOL = Pool(5)
### module2.py ###
def f(x):
# Some function
from module import POOL
POOL.map(f, range(10))
And second, if you run anything on the pool in the global scope of a module that you are importing, the system hangs. i.e. this doesn't work:
### module.py ###
from multiprocessing import Pool
def f(x): return x
p = Pool(1)
print(p.map(f, range(5)))
### module2.py ###
import module
This, however, does work, as long as nothing imports module2:
### module.py ###
from multiprocessing import Pool
def f(x): return x
p = Pool(1)
def run_pool(): print(p.map(f, range(5)))
### module2.py ###
import module
module.run_pool()
Now, the reasons behind this are only more bizarre, and likely related to the reason that the code in the question only spits an Attribute Error once each and after that appear to execute code properly. It also appears that pool threads (at least with some reliability) reload the code in module after executing.
The function you want to execute on a thread pool must be already defined when you create the pool.
This should work:
from multiprocessing import Pool
def f(x): print(x)
if __name__ == '__main__':
p = Pool(3)
p.map(f, range(20))
The reason is that (at least on Unix-based systems, which have fork) when you create a pool the workers are created by forking the current process. So if the target function isn't already defined at that point, the worker won't be able to call it.
On Windows it's a bit different, as Windows doesn't have fork. Here new worker processes are started and the main module is imported. That's why on Windows it's important to protect the executing code with a if __name__ == '__main__'. Otherwise each new worker will re-execute the code and therefore spawn new processes infinitely, crashing the program (or the system).
There is another possible source for this error. I got this error when running the example code.
The source was that despite having installed multiprosessing correctly, the C++ compiler was not installed on my system, something pip informed me of when trying to update multiprocessing. So It might be worth checking that the compiler is installed.