Shared-memory objects in multiprocessing
Question:
Suppose I have a large in memory numpy array, I have a function func that takes in this giant array as input (together with some other parameters). func with different parameters can be run in parallel. For example:
def func(arr, param):
# do stuff to arr, param
# build array arr
pool = Pool(processes = 6)
results = [pool.apply_async(func, [arr, param]) for param in all_params]
output = [res.get() for res in results]
If I use multiprocessing library, then that giant array will be copied for multiple times into different processes.
Is there a way to let different processes share the same array? This array object is read-only and will never be modified.
What’s more complicated, if arr is not an array, but an arbitrary python object, is there a way to share it?
[EDITED]
I read the answer but I am still a bit confused. Since fork() is copy-on-write, we should not invoke any additional cost when spawning new processes in python multiprocessing library. But the following code suggests there is a huge overhead:
from multiprocessing import Pool, Manager
import numpy as np;
import time
def f(arr):
return len(arr)
t = time.time()
arr = np.arange(10000000)
print "construct array = ", time.time() - t;
pool = Pool(processes = 6)
t = time.time()
res = pool.apply_async(f, [arr,])
res.get()
print "multiprocessing overhead = ", time.time() - t;
output (and by the way, the cost increases as the size of the array increases, so I suspect there is still overhead related to memory copying):
construct array = 0.0178790092468
multiprocessing overhead = 0.252444982529
Why is there such huge overhead, if we didn’t copy the array? And what part does the shared memory save me?
Answers:
If you use an operating system that uses copy-on-write fork() semantics (like any common unix), then as long as you never alter your data structure it will be available to all child processes without taking up additional memory. You will not have to do anything special (except make absolutely sure you don’t alter the object).
The most efficient thing you can do for your problem would be to pack your array into an efficient array structure (using numpy or array), place that in shared memory, wrap it with multiprocessing.Array, and pass that to your functions. This answer shows how to do that.
If you want a writeable shared object, then you will need to wrap it with some kind of synchronization or locking. multiprocessing provides two methods of doing this: one using shared memory (suitable for simple values, arrays, or ctypes) or a Manager proxy, where one process holds the memory and a manager arbitrates access to it from other processes (even over a network).
The Manager approach can be used with arbitrary Python objects, but will be slower than the equivalent using shared memory because the objects need to be serialized/deserialized and sent between processes.
There are a wealth of parallel processing libraries and approaches available in Python. multiprocessing is an excellent and well rounded library, but if you have special needs perhaps one of the other approaches may be better.
I run into the same problem and wrote a little shared-memory utility class to work around it.
I’m using multiprocessing.RawArray (lockfree), and also the access to the arrays is not synchronized at all (lockfree), be careful not to shoot your own feet.
With the solution I get speedups by a factor of approx 3 on a quad-core i7.
Here’s the code:
Feel free to use and improve it, and please report back any bugs.
'''
Created on 14.05.2013
@author: martin
'''
import multiprocessing
import ctypes
import numpy as np
class SharedNumpyMemManagerError(Exception):
pass
'''
Singleton Pattern
'''
class SharedNumpyMemManager:
_initSize = 1024
_instance = None
def __new__(cls, *args, **kwargs):
if not cls._instance:
cls._instance = super(SharedNumpyMemManager, cls).__new__(
cls, *args, **kwargs)
return cls._instance
def __init__(self):
self.lock = multiprocessing.Lock()
self.cur = 0
self.cnt = 0
self.shared_arrays = [None] * SharedNumpyMemManager._initSize
def __createArray(self, dimensions, ctype=ctypes.c_double):
self.lock.acquire()
# double size if necessary
if (self.cnt >= len(self.shared_arrays)):
self.shared_arrays = self.shared_arrays + [None] * len(self.shared_arrays)
# next handle
self.__getNextFreeHdl()
# create array in shared memory segment
shared_array_base = multiprocessing.RawArray(ctype, np.prod(dimensions))
# convert to numpy array vie ctypeslib
self.shared_arrays[self.cur] = np.ctypeslib.as_array(shared_array_base)
# do a reshape for correct dimensions
# Returns a masked array containing the same data, but with a new shape.
# The result is a view on the original array
self.shared_arrays[self.cur] = self.shared_arrays[self.cnt].reshape(dimensions)
# update cnt
self.cnt += 1
self.lock.release()
# return handle to the shared memory numpy array
return self.cur
def __getNextFreeHdl(self):
orgCur = self.cur
while self.shared_arrays[self.cur] is not None:
self.cur = (self.cur + 1) % len(self.shared_arrays)
if orgCur == self.cur:
raise SharedNumpyMemManagerError('Max Number of Shared Numpy Arrays Exceeded!')
def __freeArray(self, hdl):
self.lock.acquire()
# set reference to None
if self.shared_arrays[hdl] is not None: # consider multiple calls to free
self.shared_arrays[hdl] = None
self.cnt -= 1
self.lock.release()
def __getArray(self, i):
return self.shared_arrays[i]
@staticmethod
def getInstance():
if not SharedNumpyMemManager._instance:
SharedNumpyMemManager._instance = SharedNumpyMemManager()
return SharedNumpyMemManager._instance
@staticmethod
def createArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__createArray(*args, **kwargs)
@staticmethod
def getArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__getArray(*args, **kwargs)
@staticmethod
def freeArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__freeArray(*args, **kwargs)
# Init Singleton on module load
SharedNumpyMemManager.getInstance()
if __name__ == '__main__':
import timeit
N_PROC = 8
INNER_LOOP = 10000
N = 1000
def propagate(t):
i, shm_hdl, evidence = t
a = SharedNumpyMemManager.getArray(shm_hdl)
for j in range(INNER_LOOP):
a[i] = i
class Parallel_Dummy_PF:
def __init__(self, N):
self.N = N
self.arrayHdl = SharedNumpyMemManager.createArray(self.N, ctype=ctypes.c_double)
self.pool = multiprocessing.Pool(processes=N_PROC)
def update_par(self, evidence):
self.pool.map(propagate, zip(range(self.N), [self.arrayHdl] * self.N, [evidence] * self.N))
def update_seq(self, evidence):
for i in range(self.N):
propagate((i, self.arrayHdl, evidence))
def getArray(self):
return SharedNumpyMemManager.getArray(self.arrayHdl)
def parallelExec():
pf = Parallel_Dummy_PF(N)
print(pf.getArray())
pf.update_par(5)
print(pf.getArray())
def sequentialExec():
pf = Parallel_Dummy_PF(N)
print(pf.getArray())
pf.update_seq(5)
print(pf.getArray())
t1 = timeit.Timer("sequentialExec()", "from __main__ import sequentialExec")
t2 = timeit.Timer("parallelExec()", "from __main__ import parallelExec")
print("Sequential: ", t1.timeit(number=1))
print("Parallel: ", t2.timeit(number=1))
This is the intended use case for Ray, which is a library for parallel and distributed Python. Under the hood, it serializes objects using the Apache Arrow data layout (which is a zero-copy format) and stores them in a shared-memory object store so they can be accessed by multiple processes without creating copies.
The code would look like the following.
import numpy as np
import ray
ray.init()
@ray.remote
def func(array, param):
# Do stuff.
return 1
array = np.ones(10**6)
# Store the array in the shared memory object store once
# so it is not copied multiple times.
array_id = ray.put(array)
result_ids = [func.remote(array_id, i) for i in range(4)]
output = ray.get(result_ids)
If you don’t call ray.put then the array will still be stored in shared memory, but that will be done once per invocation of func, which is not what you want.
Note that this will work not only for arrays but also for objects that contain arrays, e.g., dictionaries mapping ints to arrays as below.
You can compare the performance of serialization in Ray versus pickle by running the following in IPython.
import numpy as np
import pickle
import ray
ray.init()
x = {i: np.ones(10**7) for i in range(20)}
# Time Ray.
%time x_id = ray.put(x) # 2.4s
%time new_x = ray.get(x_id) # 0.00073s
# Time pickle.
%time serialized = pickle.dumps(x) # 2.6s
%time deserialized = pickle.loads(serialized) # 1.9s
Serialization with Ray is only slightly faster than pickle, but deserialization is 1000x faster because of the use of shared memory (this number will of course depend on the object).
See the Ray documentation. You can read more about fast serialization using Ray and Arrow. Note I’m one of the Ray developers.
Like Robert Nishihara mentioned, Apache Arrow makes this easy, specifically with the Plasma in-memory object store, which is what Ray is built on.
I made brain-plasma specifically for this reason – fast loading and reloading of big objects in a Flask app. It’s a shared-memory object namespace for Apache Arrow-serializable objects, including pickle‘d bytestrings generated by pickle.dumps(...).
The key difference with Apache Ray and Plasma is that it keeps track of object IDs for you. Any processes or threads or programs that are running on locally can share the variables’ values by calling the name from any Brain object.
$ pip install brain-plasma
$ plasma_store -m 10000000 -s /tmp/plasma
from brain_plasma import Brain
brain = Brain(path='/tmp/plasma/')
brain['a'] = [1]*10000
brain['a']
# >>> [1,1,1,1,...]
Suppose I have a large in memory numpy array, I have a function func that takes in this giant array as input (together with some other parameters). func with different parameters can be run in parallel. For example:
def func(arr, param):
# do stuff to arr, param
# build array arr
pool = Pool(processes = 6)
results = [pool.apply_async(func, [arr, param]) for param in all_params]
output = [res.get() for res in results]
If I use multiprocessing library, then that giant array will be copied for multiple times into different processes.
Is there a way to let different processes share the same array? This array object is read-only and will never be modified.
What’s more complicated, if arr is not an array, but an arbitrary python object, is there a way to share it?
[EDITED]
I read the answer but I am still a bit confused. Since fork() is copy-on-write, we should not invoke any additional cost when spawning new processes in python multiprocessing library. But the following code suggests there is a huge overhead:
from multiprocessing import Pool, Manager
import numpy as np;
import time
def f(arr):
return len(arr)
t = time.time()
arr = np.arange(10000000)
print "construct array = ", time.time() - t;
pool = Pool(processes = 6)
t = time.time()
res = pool.apply_async(f, [arr,])
res.get()
print "multiprocessing overhead = ", time.time() - t;
output (and by the way, the cost increases as the size of the array increases, so I suspect there is still overhead related to memory copying):
construct array = 0.0178790092468
multiprocessing overhead = 0.252444982529
Why is there such huge overhead, if we didn’t copy the array? And what part does the shared memory save me?
If you use an operating system that uses copy-on-write fork() semantics (like any common unix), then as long as you never alter your data structure it will be available to all child processes without taking up additional memory. You will not have to do anything special (except make absolutely sure you don’t alter the object).
The most efficient thing you can do for your problem would be to pack your array into an efficient array structure (using numpy or array), place that in shared memory, wrap it with multiprocessing.Array, and pass that to your functions. This answer shows how to do that.
If you want a writeable shared object, then you will need to wrap it with some kind of synchronization or locking. multiprocessing provides two methods of doing this: one using shared memory (suitable for simple values, arrays, or ctypes) or a Manager proxy, where one process holds the memory and a manager arbitrates access to it from other processes (even over a network).
The Manager approach can be used with arbitrary Python objects, but will be slower than the equivalent using shared memory because the objects need to be serialized/deserialized and sent between processes.
There are a wealth of parallel processing libraries and approaches available in Python. multiprocessing is an excellent and well rounded library, but if you have special needs perhaps one of the other approaches may be better.
I run into the same problem and wrote a little shared-memory utility class to work around it.
I’m using multiprocessing.RawArray (lockfree), and also the access to the arrays is not synchronized at all (lockfree), be careful not to shoot your own feet.
With the solution I get speedups by a factor of approx 3 on a quad-core i7.
Here’s the code:
Feel free to use and improve it, and please report back any bugs.
'''
Created on 14.05.2013
@author: martin
'''
import multiprocessing
import ctypes
import numpy as np
class SharedNumpyMemManagerError(Exception):
pass
'''
Singleton Pattern
'''
class SharedNumpyMemManager:
_initSize = 1024
_instance = None
def __new__(cls, *args, **kwargs):
if not cls._instance:
cls._instance = super(SharedNumpyMemManager, cls).__new__(
cls, *args, **kwargs)
return cls._instance
def __init__(self):
self.lock = multiprocessing.Lock()
self.cur = 0
self.cnt = 0
self.shared_arrays = [None] * SharedNumpyMemManager._initSize
def __createArray(self, dimensions, ctype=ctypes.c_double):
self.lock.acquire()
# double size if necessary
if (self.cnt >= len(self.shared_arrays)):
self.shared_arrays = self.shared_arrays + [None] * len(self.shared_arrays)
# next handle
self.__getNextFreeHdl()
# create array in shared memory segment
shared_array_base = multiprocessing.RawArray(ctype, np.prod(dimensions))
# convert to numpy array vie ctypeslib
self.shared_arrays[self.cur] = np.ctypeslib.as_array(shared_array_base)
# do a reshape for correct dimensions
# Returns a masked array containing the same data, but with a new shape.
# The result is a view on the original array
self.shared_arrays[self.cur] = self.shared_arrays[self.cnt].reshape(dimensions)
# update cnt
self.cnt += 1
self.lock.release()
# return handle to the shared memory numpy array
return self.cur
def __getNextFreeHdl(self):
orgCur = self.cur
while self.shared_arrays[self.cur] is not None:
self.cur = (self.cur + 1) % len(self.shared_arrays)
if orgCur == self.cur:
raise SharedNumpyMemManagerError('Max Number of Shared Numpy Arrays Exceeded!')
def __freeArray(self, hdl):
self.lock.acquire()
# set reference to None
if self.shared_arrays[hdl] is not None: # consider multiple calls to free
self.shared_arrays[hdl] = None
self.cnt -= 1
self.lock.release()
def __getArray(self, i):
return self.shared_arrays[i]
@staticmethod
def getInstance():
if not SharedNumpyMemManager._instance:
SharedNumpyMemManager._instance = SharedNumpyMemManager()
return SharedNumpyMemManager._instance
@staticmethod
def createArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__createArray(*args, **kwargs)
@staticmethod
def getArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__getArray(*args, **kwargs)
@staticmethod
def freeArray(*args, **kwargs):
return SharedNumpyMemManager.getInstance().__freeArray(*args, **kwargs)
# Init Singleton on module load
SharedNumpyMemManager.getInstance()
if __name__ == '__main__':
import timeit
N_PROC = 8
INNER_LOOP = 10000
N = 1000
def propagate(t):
i, shm_hdl, evidence = t
a = SharedNumpyMemManager.getArray(shm_hdl)
for j in range(INNER_LOOP):
a[i] = i
class Parallel_Dummy_PF:
def __init__(self, N):
self.N = N
self.arrayHdl = SharedNumpyMemManager.createArray(self.N, ctype=ctypes.c_double)
self.pool = multiprocessing.Pool(processes=N_PROC)
def update_par(self, evidence):
self.pool.map(propagate, zip(range(self.N), [self.arrayHdl] * self.N, [evidence] * self.N))
def update_seq(self, evidence):
for i in range(self.N):
propagate((i, self.arrayHdl, evidence))
def getArray(self):
return SharedNumpyMemManager.getArray(self.arrayHdl)
def parallelExec():
pf = Parallel_Dummy_PF(N)
print(pf.getArray())
pf.update_par(5)
print(pf.getArray())
def sequentialExec():
pf = Parallel_Dummy_PF(N)
print(pf.getArray())
pf.update_seq(5)
print(pf.getArray())
t1 = timeit.Timer("sequentialExec()", "from __main__ import sequentialExec")
t2 = timeit.Timer("parallelExec()", "from __main__ import parallelExec")
print("Sequential: ", t1.timeit(number=1))
print("Parallel: ", t2.timeit(number=1))
This is the intended use case for Ray, which is a library for parallel and distributed Python. Under the hood, it serializes objects using the Apache Arrow data layout (which is a zero-copy format) and stores them in a shared-memory object store so they can be accessed by multiple processes without creating copies.
The code would look like the following.
import numpy as np
import ray
ray.init()
@ray.remote
def func(array, param):
# Do stuff.
return 1
array = np.ones(10**6)
# Store the array in the shared memory object store once
# so it is not copied multiple times.
array_id = ray.put(array)
result_ids = [func.remote(array_id, i) for i in range(4)]
output = ray.get(result_ids)
If you don’t call ray.put then the array will still be stored in shared memory, but that will be done once per invocation of func, which is not what you want.
Note that this will work not only for arrays but also for objects that contain arrays, e.g., dictionaries mapping ints to arrays as below.
You can compare the performance of serialization in Ray versus pickle by running the following in IPython.
import numpy as np
import pickle
import ray
ray.init()
x = {i: np.ones(10**7) for i in range(20)}
# Time Ray.
%time x_id = ray.put(x) # 2.4s
%time new_x = ray.get(x_id) # 0.00073s
# Time pickle.
%time serialized = pickle.dumps(x) # 2.6s
%time deserialized = pickle.loads(serialized) # 1.9s
Serialization with Ray is only slightly faster than pickle, but deserialization is 1000x faster because of the use of shared memory (this number will of course depend on the object).
See the Ray documentation. You can read more about fast serialization using Ray and Arrow. Note I’m one of the Ray developers.
Like Robert Nishihara mentioned, Apache Arrow makes this easy, specifically with the Plasma in-memory object store, which is what Ray is built on.
I made brain-plasma specifically for this reason – fast loading and reloading of big objects in a Flask app. It’s a shared-memory object namespace for Apache Arrow-serializable objects, including pickle‘d bytestrings generated by pickle.dumps(...).
The key difference with Apache Ray and Plasma is that it keeps track of object IDs for you. Any processes or threads or programs that are running on locally can share the variables’ values by calling the name from any Brain object.
$ pip install brain-plasma
$ plasma_store -m 10000000 -s /tmp/plasma
from brain_plasma import Brain
brain = Brain(path='/tmp/plasma/')
brain['a'] = [1]*10000
brain['a']
# >>> [1,1,1,1,...]