We Keep Coding

Simple way to parallelize embarrassingly parallelizable generator

I have a generator (or, a list of generators). Let's call them gens
Each generator in gens is a complicated function that returns the next value of a complicated procedure. Fortunately, they are all independent of one another.
I want to call gen.__next__() for each element gen in gens, and return the resulting values in a list. However, multiprocessing is unhappy with pickling generators.
Is there a fast, simple way to do this in Python? I would like it such that gens of length m is mapped to n cores locally on my machine, where n could be larger or smaller than m. Each generator should run on a separate core.
If this is possible, can someone provide a minimal example?

You can't pickle generators. Read more about it here.
There is a blog post which explains it in much more detail. Referring a quote from it:
Let’s ignore that problem for a moment and look what we would need to do to pickle a generator. Since a generator is essentially a souped-up function, we would need to save its bytecode, which is not guarantee to be backward-compatible between Python’s versions, and its frame, which holds the state of the generator such as local variables, closures and the instruction pointer. And this latter is rather cumbersome to accomplish, since it basically requires to make the whole interpreter picklable. So, any support for pickling generators would require a large number of changes to CPython’s core.
Now if an object unsupported by pickle (e.g., a file handle, a socket, a database connection, etc) occurs in the local variables of a generator, then that generator could not be pickled automatically, regardless of any pickle support for generators we might implement. So in that case, you would still need to provide custom getstate and setstate methods. This problem renders any pickling support for generators rather limited.
He also suggests a solution, to use simple iterators.
the best solution to this problem to the rewrite the generators as simple iterators (i.e., one with a __next__ method). Iterators are easy and efficient space-wise to pickle because their state is explicit. You would still need to handle objects representing some external state explicitly however; you cannot get around this.
Another offered solution (which I haven't tried) suggests this
Con­vert the gen­er­a­tor to a class in which the gen­er­a­tor code is the __iter__ method
Add __getstate__ and __setstate__ meth­ods to the class, to han­dling pick­ling. Remem­ber that you can’t pickle file objects. So __setstate__ will have to re-open files, as necessary.

If your subtasks are truly parallel (do not rely on any shared state), you can do this with multiprocesing.Pool().
Take a look at
https://docs.python.org/3/library/multiprocessing.html#multiprocessing.pool.Pool
This requires you to make the arguments of pool.map() serializable. You can't pass in a generator to your worker, but you can achieve something similar by defining your generator inside the target function, and pass in initialization arguments to the multiprocessing library:
import multiprocessing as mp
import time
def worker(value):
# The generator is defined inside the multiprocessed function
def gen():
for k in range(value):
time.sleep(1) # Simulate long running task
yield k
# Execute the generator
for x in gen():
print(x)
# Do something with x?
pass
pool = mp.Pool()
pool.map(worker, [2, 5, 2])
pool.join() # Wait for all the work to be finished.
pool.close() # Clean up system resources
The output will be:
0
0
0
1
1
1
2
3
4
Note that this solution only really works if you build your generators, then use them only once, as their final state is lost at the end of the worker function.
Keep in mind that anytime you want to use multiprocessing, you have to use for serializable objects due to the limitations of inter-process communication; this can often prove limiting.
If your process is not CPU bound but instead I/O bound (disk access, network access, etc), you'll have a much easier time using threads.

You don't need to pickle the generator, just send an index of the generator to the processing pool.
M = len(gens)
N = multiprocessing.cpu_count()
def proc(gen_idx):
return [r for r in gens[gen_idx]()]
if __name__ == "__main__":
with multiprocessing.Pool(N) as p:
for r in p.imap_unordered(proc, range(M)):
print(r)
Note that I don't call/initialize the generator until within the processing function.
Using imap_unordered will allow you to process the results as each generator completes.

It's quite easy to implement, just dont block the threads sincronusly, just constantly loop thru the states and join them on complition. This template shuld be good enough to give an idea, self.done alwais needs to be set last on thread complition and las on thread reuse.
import threading as th
import random
import time
class Gen_thread(th.Thread):
def is_done(self):
return self.done
def get_result(self):
return self.work_result
def __init__(self, *args, **kwargs):
self.g_id = kwargs['id']
self.kwargs = kwargs
self.args = args
self.work_result = None
self.done = False
th.Thread.__init__(self)
def run(self):
# time.sleep(*self.args) to pass variables
time.sleep(random.randint(1, 4))
self.work_result = 'Thread {0} done'.format(self.g_id + 1)
self.done = True
class Gens(object):
def __init__(self, n):
self.n_needed = 0
self.n_done = 0
self.n_loop = n
self.workers_tmp = None
self.workers = []
def __iter__(self):
return self
def __next__(self):
if self.n_needed == 0:
for w in range(self.n_loop):
self.workers.append(Gen_thread(id=w))
self.workers[w].start()
self.n_needed += 1
while self.n_done != self.n_needed:
for w in range(self.n_loop):
if self.workers[w].is_done():
self.workers[w].join()
self.workers_tmp = self.workers[w].get_result()
self.workers.pop(w)
self.n_loop -= 1
self.n_done += 1
return self.workers_tmp
raise StopIteration()
if __name__ == '__main__':
for gen in Gens(4):
print(gen)

how to combine dask and classes?

I am trying to rewrite an entire project that has been developed with classes. Little by little, the heaviest computational chunks should be parallelized, clearly we have a lot of independent sequential loops. An example with classes that mimicks the behaviour is this toy problem (I'm a mathematician obsessed with the p-sums):
class Summer:
def __init__(self, p):
self.p = p
def sum(self):
return sum(pow(i,-self.p) for i in range(1,1000000))
total = sum([Summer(p).sum() for p in range(2,20)])
If I replace the last line by:
from dask.distributed import Client
def psum(p): return Summer(p).sum()
client = Client()
A = client.map(psum,range(2,20))
total=client.submit(sum,A).result()
My runtime is cut in 4 (the number of cores available on my machine). This ideal behaviour does NOT persist if I use my real classes which are data intensive (big pandas structures taking up memory). Is there a recommended alternative to dask.distributed? I'm seeing bad slowdowns which I attribute to data being passed around.

Python Multiprocessing Slower and not really working for object methods

Edit: Running Apple MBP 2017 Model 14,3 with 2.8GHz i7 4-cores:
multiprocessing.cpu_count()
8
I have a list of objects I'm performing object methods on in python once for each object. The process is for a genetic algorithm so I'm interested in speeding it up. Basically, each time I update the environment with data from the data list, the object (genome) performs a little bit of math including taking values from the environment, and referencing it's own internal values.
I'm doing:
from multiprocessing import Pool
class Individual(object):
def __init__(self):
self.parameter1 = None
self.parameter2 = None
def update_values():
# reads the environment variables, does math specific to each instance
# updates internal parameters
a, b, c, d = environment_variables
self.parameter1 = do_math(a, b, c, d,
self.parameter1, self.parameter2)
self.parameter2 = do_math(a, b, c, d,
self.parameter1, self.parameter2)
data_list = [data1, data2, data3, ..., data1000]
object_list = [object1, object2, object3, ..., object20000]
If I run this:
for newdataset in data_list:
update_parameters(newdataset)
for object in object_list:
object.update_values()
It is much faster than if I try to split this up using multiprocessing/ map:
def process_object(object):
object.update_values()
for newdataset in data_list:
update_parameters(newdataset)
with Pool(4) as p:
p.map(process_object, object_list)
If I run with object_list length of 200 (instead of 20000) the total time is 14.8 seconds in single threaded mode.
If I run with the same in multiprocessing mode the total time is... still waiting... ok... 211 seconds.
Also it doesn't appear to do what the function says it should at all. What am I missing here? When I check the values of each object they do not appear to have been updated at all.

When you use multiprocessing, you're serializing and transferring the data both ways. In this case, that includes each object you indend to call update_values on. I'm guessing that you're also iterating on your models, meaning they'll be sent back and forth quite a lot. Furthermore, map() returns a list of results, but process_object just returns None. So you've serialized a model, sent it to another process, had that process run and update the model, then send a None back and toss away the updated model, before tossing away the list of None results. If you were to return the models:
def process_object(object):
object.update_values()
return object
...
object_list = p.map(process_object, object_list)
Your program might actually produce some results, but almost certainly still slower than you wish. In particular your process pool will not have the data_list or similar things (the "environment"?) - it only receives what you passed through Pool.map().
You may want to consider using other tools such as tensorflow or MPI. At least read up on sharing state between processes. Also, you probably shouldn't be recreating your process pool for every iteration; that's very expensive on some platforms, such as Windows.

I would split up the parallelization a little bit differently. It's hard to tell what's happening with update_parameters, but I would parallelize the call to that too. Why leave it out? You could wrap the whole operation you're interested in, in some function, right?
Also, this is important: you need to make sure that you only open up the pool if you're in the main process. So add the line
if __name__ == '__main__':
with Pool(multiprocessing.cpu_count()) as p:

Is filter thread-safe

I have a thread which is updating a list called l. Am I right in saying that it is thread-safe to do the following from another thread?
filter(lambda x: x[0] == "in", l)
If its not thread safe, is this then the correct approach:
import threading
import time
import Queue
class Logger(threading.Thread):
def __init__(self, log):
super(Logger, self).__init__()
self.log = log
self.data = []
self.finished = False
self.data_lock = threading.Lock()
def run(self):
while not self.finished:
try:
with self.data_lock:
self.data.append(self.log.get(block=True, timeout=0.1))
except Queue.Empty:
pass
def get_data(self, cond):
with self.data_lock:
d = filter(cond, self.data)
return d
def stop(self):
self.finished = True
self.join()
print("Logger stopped")
where the get_data(self, cond) method is used to retrieve a small subset of the data in the self.data in a thread safe manner.

First, to answer your question in the title: filter is just a function. Hence, its thread-safety will rely on the data-structure you use it with.
As pointed out in the comments already, list operations themselves are thread-safe in CPython and protected by the GIL, but that is arguably only an implementation detail of CPython that you shouldn't really rely on. Even if you could rely on it, thread safety of some of their operations probably does not mean the kind of thread safety you mean:
The problem is that iterating over a sequence with filter is in general not an atomic operation. The sequence could be changed during iteration. Depending on the data-structure underlying your iterator this might cause more or less weird effects. One way to overcome this problem is by iterating over a copy of the sequence that is created with one atomic action. Easiest way to do this for standard sequences like tuple, list, string is with the slice operator like this:
filter(lambda x: x[0] == "in", l[:])
Apart from this not necessarily being thread-safe for other data-types, there's one problem with this though: it's only a shallow copy. As your list's elements seem to be list-like as well, another thread could in parallel do del l[1000][:] to empty one of the inner lists (which are pointed to in your shallow copy as well). This would make your filter expression fail with an IndexError.
All that said, it's not a shame to use a lock to protect access to your list and I'd definitely recommend it. Depending on how your data changes and how you work with the returned data, it might even be wise to deep-copy the elements while holding the lock and to return those copies. That way you can guarantee that once returned the filter condition won't suddenly change for the returned elements.
Wrt. your Logger code: I'm not 100 % sure how you plan to use this and if it's critical for you to run several threads on one queue and join them. What looks weird to me is that you never use Queue.task_done() (assuming that its self.log is a Queue). Also your polling of the queue is potentially wasteful. If you don't need the join of the thread, I'd suggest to at least turn the lock acquisition around:
class Logger(threading.Thread):
def __init__(self, log):
super(Logger, self).__init__()
self.daemon = True
self.log = log
self.data = []
self.data_lock = threading.Lock()
def run(self):
while True:
l = self.log.get() # thread will sleep here indefinitely
with self.data_lock:
self.data.append(l)
self.log.task_done()
def get_data(self, cond):
with self.data_lock:
d = filter(cond, self.data)
# maybe deepcopy d here
return d
Externally you could still do log.join() to make sure that all of the elements of the log queue are processed.

If one thread writes to a list and another thread reads that list, the two must be synchronized. It doesn't matter for that aspect whether the reader uses filter(), an index or iteration or whether the writer uses append() or any other method.
In your code, you achieve the necessary synchronization using a threading.Lock. Since you only access the list within the context of with self.data_lock, the accesses are mutually exclusive.
In summary, your code is formally correct concerning the list handling between threads. But:
You do access self.finished without the lock, which is problematic. Assigning to that member will change self, i.e. the mapping of the object to the according members, so this should be synced. Effectively, this won't hurt, because True and False are global constants, at the worst you will have a short delay between setting the state in one thread and seeing the state in the other. It remains bad, because it is habit-forming.
As a rule, when you use a lock, always document which objects this lock protects. Also, document which object is accessed by which thread. The fact that self.finished is shared and requires synchronization would have been obvious. Also, making a visual distinction between public functions and data and private ones (beginning with an _underscore, see PEP 8) helps keeping track of this. It also helps other readers.
A similar issue is your baseclass. In general, inheriting from threading.Thread is a bad idea. Rather, include an instance of the thread class and give it a function like self._main_loop to run on. The reason is that you say that your Logger is a Thread and that all of its baseclass' public members are also public members of your class, which is probably a much wider interface than what you intended.
You should never block with a lock held. In your code, you block in self.log.get(block=True, timeout=0.1) with the lock on the mutex. In that time, even if nothing actually happens, no other thread will be able to call and complete a call to get_data(). There is actually just a tiny window between unlocking the mutex and locking it again where a caller of get_data() does not have to wait, which is very bad for performance. I could even imagine that your question is motivated by the really bad performance this causes. Instead, call log.get(..) without lock, it shouldn't need one. Then, with the lock held, append data to self.data and check self.finished.

Python multiprocess share memory vs using arguments

I'm trying to get my head around what is the most efficient and less memory consuming way to share the same data source between different process.
Imagine the following code, that simplify my problem.
import pandas as pd
import numpy as np
from multiprocessing import Pool
# method #1
def foo(i): return data[i]
if __name__ == '__main__':
data = pd.Series(np.array(range(100000)))
pool = Pool(2)
print pool.map(foo,[10,134,8,1])
# method #2
def foo((data,i)): return data[i]
if __name__ == '__main__':
data = pd.Series(np.array(range(100000)))
pool = Pool(2)
print pool.map(foo,[(data,10),(data,134),(data,8),(data,1)])
In the first method will use the global variable (won't work on Windows, only on Linux/OSX) which will then access by the function. In the second method I'm passing "data" as part of the arguments.
In terms of memory used during the process, there will be a difference between the two methods?
# method #3
def foo((data,i)): return data[i]
if __name__ == '__main__':
data = pd.Series(np.array(range(100000)))
pool = Pool(2)
# reduce the size of the argument passed
data1 = data[:1000]
print pool.map(foo,[(data1,10),(data1,134),(data1,8),(data1,1)])
A third method, rather than passing all the "data", since we know we'll be using only the first records, I'm only passing the first 1000 records. Will this make any difference?
Background
The problem I'm facing I have a big dataset of about 2 millions rows (4GB in memory) which will then by four subprocess to do some elaboration. Each elaboration only affect a small portion of the data (20000 rows) and I'd like to minimize the memory use by each concurrent process.

I'm going to start with the second and third methods, because they're easier to explain.
When you pass the arguments to pool.map or pool.apply, the arguments will be pickled, sent to the child process using a pipe, and then unpickled in the child. This of course requires two completely distinct copies of the data structures you're passing. It also can lead to slow performance with large data structures, since pickling/unpickling large objects can take quite a while.
With the third method, you're just passing smaller data structures than method two. This should perform better, since you don't need to pickle/unpickle as much data.
One other note - passing data multiple times is definitely a bad idea, because each copy will be getting pickled/unpickled repeatedly. You want to pass it to each child once. Method 1 is a good way to do that, or you can use the initializer keyword argument to explicitly pass data to the child. This will use fork on Linux and pickling on Windows to pass data to the child process:
import pandas as pd
import numpy as np
from multiprocessing import Pool
data = None
def init(_data):
global data
data = _data # data is now accessible in all children, even on Windows
# method #1
def foo(i): return data[i]
if __name__ == '__main__':
data = pd.Series(np.array(range(100000)))
pool = Pool(2, initializer=init, initargs=(data,))
print pool.map(foo,[10,134,8,1])
Using the first method, you're leveraging the behavior of fork to allow the child process to inherit the data object. fork has copy-on-write semantics, which means that the memory is actually shared between the parent and its children, until you try to write to it in the child. When you try to write, the memory page that the data you're trying to write is contained in must be copied, to keep it separate from the parent version.
Now, this sounds like a slam dunk - no need to copy anything as long as we don't write to it, which is surely faster than the pickle/unpickle method. And that's usually the case. However, in practice, Python is internally writing to its objects, even when you wouldn't really expect it to. Because Python uses reference counting for memory management, it needs to increment the internal reference counter on each object every time its passed to a method, or assigned to variable, etc. So, that means the memory page containing the reference count for each object passed to your child process will end up getting copied. This will definitely be faster and use less memory than pickling data multiple times, but isn't quite completely shared, either.