Related
I am trying to run a model multiple times. As a result it is time consuming. As a solution I try to make it parallel. However, it ends up to be slower. Parallel is 40 seconds while serial is 34 seconds.
# !pip install --target=$nb_path transformers
oracle = pipeline(model="deepset/roberta-base-squad2")
question = 'When did the first extension of the Athens Tram take place?'
print(data)
print("Data size is: ", len(data))
parallel = True
if parallel == False:
counter = 0
l = len(data)
cr = []
for words in data:
counter+=1
print(counter, " out of ", l)
cr.append(oracle(question=question, context=words))
elif parallel == True:
from multiprocessing import Process, Queue
import multiprocessing
no_CPU = multiprocessing.cpu_count()
print("Number of cpu : ", no_CPU)
l = len(data)
def answer_question(data, no_CPU, sub_no):
cr_process = []
counter_process = 0
for words in data:
counter_process+=1
l_data = len(data)
# print("n is", no_CPU)
# print("l is", l_data)
print(counter_process, " out of ", l_data, "in subprocess number", sub_no)
cr_process.append(oracle(question=question, context=words))
# Q.put(cr_process)
cr.append(cr_process)
n = no_CPU # number of subprocesses
m = l//n # number of data the n-1 first subprocesses will handle
res = l % n # number of extra data samples the last subprocesses has
# print(m)
# print(res)
procs = []
# instantiating process with arguments
for x in range(n-1):
# print(x*m)
# print((x+1)*m)
proc = Process(target=answer_question, args=(data[x*m:(x+1)*m],n, x+1,))
procs.append(proc)
proc.start()
proc = Process(target=answer_question, args=(data[(n-1)*m:n*m+res],n,n,))
procs.append(proc)
proc.start()
# complete the processes
for proc in procs:
proc.join()
A sample of the data variable can be found here (to not flood the question). Argument parallel controls the serial and the parallel version. So my question is, why does it happen and how do I make the parallel version faster? I use google colab so it has 2 CPU cores available , that's what multiprocessing.cpu_count() is saying at least.
Your pipeline is already running on multi-cpu even when run as one process. The code of transformers are optimized to run on multi-cpu.
when on top of that you are creating multiple process, you are loosing some time for building the processes and switching between them.
To verify this, on the so-called "single process" version look at your cpu utilizations, you should already see all are at max, so creating extra parallel processes are not going to save you some time,
The Multiprocessing module is quite confusing for python beginners specially for those who have just migrated from MATLAB and are made lazy with its parallel computing toolbox. I have the following function which takes ~80 Secs to run and I want to shorten this time by using Multiprocessing module of Python.
from time import time
xmax = 100000000
start = time()
for x in range(xmax):
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
end = time()
tt = end-start #total time
print('Each iteration took: ', tt/xmax)
print('Total time: ', tt)
This outputs as expected:
Condition met at: -15 0
Condition met at: -3 1
Condition met at: 11 2
Each iteration took: 8.667453265190124e-07
Total time: 86.67453265190125
As any iteration of the loop is not dependent on others, I tried to adopt this Server Process from the official documentation to scan chunks of the range in separate processes. And finally I came up with vartec's answer to this question and could prepare the following code. I also updated the code based on Darkonaut's response to the current question.
from time import time
import multiprocessing as mp
def chunker (rng, t): # this functions makes t chunks out of rng
L = rng[1] - rng[0]
Lr = L % t
Lm = L // t
h = rng[0]-1
chunks = []
for i in range(0, t):
c = [h+1, h + Lm]
h += Lm
chunks.append(c)
chunks[t-1][1] += Lr + 1
return chunks
def worker(lock, xrange, return_dict):
'''worker function'''
for x in range(xrange[0], xrange[1]):
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
return_dict['x'].append(x)
return_dict['y'].append(y)
with lock:
list_x = return_dict['x']
list_y = return_dict['y']
list_x.append(x)
list_y.append(y)
return_dict['x'] = list_x
return_dict['y'] = list_y
if __name__ == '__main__':
start = time()
manager = mp.Manager()
return_dict = manager.dict()
lock = manager.Lock()
return_dict['x']=manager.list()
return_dict['y']=manager.list()
xmax = 100000000
nw = mp.cpu_count()
workers = list(range(0, nw))
chunks = chunker([0, xmax], nw)
jobs = []
for i in workers:
p = mp.Process(target=worker, args=(lock, chunks[i],return_dict))
jobs.append(p)
p.start()
for proc in jobs:
proc.join()
end = time()
tt = end-start #total time
print('Each iteration took: ', tt/xmax)
print('Total time: ', tt)
print(return_dict['x'])
print(return_dict['y'])
which considerably reduces the run time to ~17 Secs. But, my shared variable cannot retrieve any values. Please help me find out which part of the code is going wrong.
the output I get is:
Each iteration took: 1.7742713451385497e-07
Total time: 17.742713451385498
[]
[]
from which I expect:
Each iteration took: 1.7742713451385497e-07
Total time: 17.742713451385498
[0, 1, 2]
[-15, -3, 11]
The issue in your example is that modifications to standard mutable structures within Manager.dict will not be propagated. I'm first showing you how to fix it with manager, just to show you better options afterwards.
multiprocessing.Manager is a bit heavy since it uses a separate Process just for the Manager and working on a shared object needs using locks for data consistency. If you run this on one machine, there are better options with multiprocessing.Pool, in case you don't have to run customized Process classes and if you have to, multiprocessing.Process together with multiprocessing.Queue would be the common way of doing it.
The quoting parts are from the multiprocessing docs.
Manager
If standard (non-proxy) list or dict objects are contained in a referent, modifications to those mutable values will not be propagated through the manager because the proxy has no way of knowing when the values contained within are modified. However, storing a value in a container proxy (which triggers a setitem on the proxy object) does propagate through the manager and so to effectively modify such an item, one could re-assign the modified value to the container proxy...
In your case this would look like:
def worker(xrange, return_dict, lock):
"""worker function"""
for x in range(xrange[0], xrange[1]):
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
with lock:
list_x = return_dict['x']
list_y = return_dict['y']
list_x.append(x)
list_y.append(y)
return_dict['x'] = list_x
return_dict['y'] = list_y
The lock here would be a manager.Lock instance you have to pass along as argument since the whole (now) locked operation is not by itself atomic. (Here
is an easier example with Manager using Lock)
This approach is perhaps less convenient than employing nested Proxy Objects for most use cases but also demonstrates a level of control over the synchronization.
Since Python 3.6 proxy objects are nestable:
Changed in version 3.6: Shared objects are capable of being nested. For example, a shared container object such as a shared list can contain other shared objects which will all be managed and synchronized by the SyncManager.
Since Python 3.6 you can fill your manager.dict before starting multiprocessing with manager.list as values and then append directly in the worker without having to reassign.
return_dict['x'] = manager.list()
return_dict['y'] = manager.list()
EDIT:
Here is the full example with Manager:
import time
import multiprocessing as mp
from multiprocessing import Manager, Process
from contextlib import contextmanager
# mp_util.py from first link in code-snippet for "Pool"
# section below
from mp_utils import calc_batch_sizes, build_batch_ranges
# def context_timer ... see code snippet in "Pool" section below
def worker(batch_range, return_dict, lock):
"""worker function"""
for x in batch_range:
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
with lock:
return_dict['x'].append(x)
return_dict['y'].append(y)
if __name__ == '__main__':
N_WORKERS = mp.cpu_count()
X_MAX = 100000000
batch_sizes = calc_batch_sizes(X_MAX, n_workers=N_WORKERS)
batch_ranges = build_batch_ranges(batch_sizes)
print(batch_ranges)
with Manager() as manager:
lock = manager.Lock()
return_dict = manager.dict()
return_dict['x'] = manager.list()
return_dict['y'] = manager.list()
tasks = [(batch_range, return_dict, lock)
for batch_range in batch_ranges]
with context_timer():
pool = [Process(target=worker, args=args)
for args in tasks]
for p in pool:
p.start()
for p in pool:
p.join()
# Create standard container with data from manager before exiting
# the manager.
result = {k: list(v) for k, v in return_dict.items()}
print(result)
Pool
Most often a multiprocessing.Pool will just do it. You have an additional challenge in your example since you want to distribute iteration over a range.
Your chunker function doesn't manage to divide the range even so every process has about the same work to do:
chunker((0, 21), 4)
# Out: [[0, 4], [5, 9], [10, 14], [15, 21]] # 4, 4, 4, 6!
For the code below please grab the code snippet for mp_utils.py from my answer here, it provides two functions to chunk ranges as even as possible.
With multiprocessing.Pool your worker function just has to return the result and Pool will take care of transporting the result back over internal queues back to the parent process. The result will be a list, so you will have to rearange your result again in a way you want it to have. Your example could then look like this:
import time
import multiprocessing as mp
from multiprocessing import Pool
from contextlib import contextmanager
from itertools import chain
from mp_utils import calc_batch_sizes, build_batch_ranges
#contextmanager
def context_timer():
start_time = time.perf_counter()
yield
end_time = time.perf_counter()
total_time = end_time-start_time
print(f'\nEach iteration took: {total_time / X_MAX:.4f} s')
print(f'Total time: {total_time:.4f} s\n')
def worker(batch_range):
"""worker function"""
result = []
for x in batch_range:
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
result.append((x, y))
return result
if __name__ == '__main__':
N_WORKERS = mp.cpu_count()
X_MAX = 100000000
batch_sizes = calc_batch_sizes(X_MAX, n_workers=N_WORKERS)
batch_ranges = build_batch_ranges(batch_sizes)
print(batch_ranges)
with context_timer():
with Pool(N_WORKERS) as pool:
results = pool.map(worker, iterable=batch_ranges)
print(f'results: {results}')
x, y = zip(*chain.from_iterable(results)) # filter and sort results
print(f'results sorted: x: {x}, y: {y}')
Example Output:
[range(0, 12500000), range(12500000, 25000000), range(25000000, 37500000),
range(37500000, 50000000), range(50000000, 62500000), range(62500000, 75000000), range(75000000, 87500000), range(87500000, 100000000)]
Condition met at: -15 0
Condition met at: -3 1
Condition met at: 11 2
Each iteration took: 0.0000 s
Total time: 8.2408 s
results: [[(0, -15), (1, -3), (2, 11)], [], [], [], [], [], [], []]
results sorted: x: (0, 1, 2), y: (-15, -3, 11)
Process finished with exit code 0
If you had multiple arguments for your worker you would build a "tasks"-list with argument-tuples and exchange pool.map(...) with pool.starmap(...iterable=tasks). See docs for further details on that.
Process & Queue
If you can't use multiprocessing.Pool for some reason, you have to take
care of inter-process communication (IPC) yourself, by passing a
multiprocessing.Queue as argument to your worker-functions in the child-
processes and letting them enqueue their results to be send back to the
parent.
You will also have to build your Pool-like structure so you can iterate over it to start and join the processes and you have to get() the results back from the queue. More about Queue.get usage I've written up here.
A solution with this approach could look like this:
def worker(result_queue, batch_range):
"""worker function"""
result = []
for x in batch_range:
y = ((x+5)**2+x-40)
if y <= 0xf+1:
print('Condition met at: ', y, x)
result.append((x, y))
result_queue.put(result) # <--
if __name__ == '__main__':
N_WORKERS = mp.cpu_count()
X_MAX = 100000000
result_queue = mp.Queue() # <--
batch_sizes = calc_batch_sizes(X_MAX, n_workers=N_WORKERS)
batch_ranges = build_batch_ranges(batch_sizes)
print(batch_ranges)
with context_timer():
pool = [Process(target=worker, args=(result_queue, batch_range))
for batch_range in batch_ranges]
for p in pool:
p.start()
results = [result_queue.get() for _ in batch_ranges]
for p in pool:
p.join()
print(f'results: {results}')
x, y = zip(*chain.from_iterable(results)) # filter and sort results
print(f'results sorted: x: {x}, y: {y}')
I'm trying to learn how to do parallel programming in python. I wrote a simple int square function and then ran it in serial, multi-thread, and multi-process:
import time
import multiprocessing, threading
import random
def calc_square(numbers):
sq = 0
for n in numbers:
sq = n*n
def splita(list, n):
a = [[] for i in range(n)]
counter = 0
for i in range(0,len(list)):
a[counter].append(list[i])
if len(a[counter]) == len(list)/n:
counter = counter +1
continue
return a
if __name__ == "__main__":
random.seed(1)
arr = [random.randint(1, 11) for i in xrange(1000000)]
print "init completed"
start_time2 = time.time()
calc_square(arr)
end_time2 = time.time()
print "serial: " + str(end_time2 - start_time2)
newarr = splita(arr,8)
print 'split complete'
start_time = time.time()
for i in range(8):
t1 = threading.Thread(target=calc_square, args=(newarr[i],))
t1.start()
t1.join()
end_time = time.time()
print "mt: " + str(end_time - start_time)
start_time = time.time()
for i in range(8):
p1 = multiprocessing.Process(target=calc_square, args=(newarr[i],))
p1.start()
p1.join()
end_time = time.time()
print "mp: " + str(end_time - start_time)
Output:
init completed
serial: 0.0640001296997
split complete
mt: 0.0599999427795
mp: 2.97099995613
However, as you can see, something weird happened and mt is taking the same time as serial and mp is actually taking significantly longer (almost 50 times longer).
What am I doing wrong? Could someone push me in the right direction to learn parallel programming in python?
Edit 01
Looking at the comments, I see that perhaps the function not returning anything seems pointless. The reason I'm even trying this is because previously I tried the following add function:
def addi(numbers):
sq = 0
for n in numbers:
sq = sq + n
return sq
I tried returning the addition of each part to a serial number adder, so at least I could see some performance improvement over a pure serial implementation. However, I couldn't figure out how to store and use the returned value, and that's the reason I'm trying to figure out something even simpler than that, which is just dividing up the array and running a simple function on it.
Thanks!
I think that multiprocessing takes quite a long time to create and start each process. I have changed the program to make 10 times the size of arr and changed the way that the processes are started and there is a slight speed-up:
(Also note python 3)
import time
import multiprocessing, threading
from multiprocessing import Queue
import random
def calc_square_q(numbers,q):
while q.empty():
pass
return calc_square(numbers)
if __name__ == "__main__":
random.seed(1) # note how big arr is now vvvvvvv
arr = [random.randint(1, 11) for i in range(10000000)]
print("init completed")
# ...
# other stuff as before
# ...
processes=[]
q=Queue()
for arrs in newarr:
processes.append(multiprocessing.Process(target=calc_square_q, args=(arrs,q)))
print('start processes')
for p in processes:
p.start() # even tho' each process is started it waits...
print('join processes')
q.put(None) # ... for q to become not empty.
start_time = time.time()
for p in processes:
p.join()
end_time = time.time()
print("mp: " + str(end_time - start_time))
Also notice above how I create and start the processes in two different loops, and then finally join with the processes in a third loop.
Output:
init completed
serial: 0.53214430809021
split complete
start threads
mt: 0.5551605224609375
start processes
join processes
mp: 0.2800724506378174
Another factor of 10 increase in size of arr:
init completed
serial: 5.8455305099487305
split complete
start threads
mt: 5.411392450332642
start processes
join processes
mp: 1.9705185890197754
And yes, I've also tried this in python 2.7, although Threads seemed slower.
I want to parallelize a task using python, so I read about pool.map, where data is divided into multiple chunks and processed by each process (thread).
I have a huge dictionary(2 million words) and a text file of sentences, the idea is to divide sentences into words and match each word to the exiting dictionary and do further processing based on the return result. Before doing that ,I wrote a dummy program to check the functionality of pool.map but it is not working as expected (i.e single process takes less time than multiple process) (I am using process and thread interchangeably because I think every thread is nothing but a process here)
def add_1(x):
return (x*x+x)
def main():
iter = 10000000
num = [i for i in xrange(iter)]
threads = 4
pool = ThreadPool(threads)
start = time.time()
results = pool.map(add_1,num,iter/threads)
pool.close()
pool.join()
end = time.time()
print('Total Time Taken = %f')% (end-start)
Total Time Taken:
Thread 1 Total Time Taken = 2.252361
Thread 2 Total Time Taken = 2.560798
Thread 3 Total Time Taken = 2.938640
Thread 4 Total Time Taken = 3.048179
Just using pool = ThreadPool()
def main:
num = [i for i in xrange(iter)]
#pool = ThreadPool(threads)
pool = ThreadPool()
start = time.time()
#results = pool.map(add_1,num,iter/threads)
results = pool.map(add_1,num)
pool.close()
pool.join()
end = time.time()
print('Total Time Taken = %f')% (end-start)
Total Time Taken = 3.031125
Normal for loop execution:
def main():
iter = 10000000
start = time.time()
for k in xrange(iter):
add_1(k)
end = time.time()
print ('Total Time normally = %f') % (end-start)
Total Time normally = 1.087591
Config:
I am using python 2.7.6
I am trying to make a multiprocessing MongoDB utility, it is perfectly working, but I think I have a performance issue... Even with 20 workers,it isn't processing more than 2800 docs per second... I think I can get 5x faster... This is my code, it isn't doing anything exceptional, just prints a remaining time to the end of the cursor.
Maybe there is a better way to perform multiprocessing on a MongoDB cursor, because I need to run some stuff on every doc with a 17.4M records collection, so performance and less time is a must.
START = time.time()
def remaining_time(a, b):
if START:
y = (time.time() - START)
z = ((a * y) / b) - y
d = time.strftime('%H:%M:%S', time.gmtime(z))
e = round(b / y)
progress("{0}/{1} | Tiempo restante {2} ({3}p/s)".format(b, a, d, e), b, a)
def progress(p, c, t):
pc = (c * 100) / t
sys.stdout.write("%s [%-20s] %d%%\r" % (p, '█' * (pc / 5), pc))
sys.stdout.flush()
def dowork(queue):
for p, i, pcount in iter(queue.get, 'STOP'):
remaining_time(pcount, i)
def populate_jobs(queue):
mongo_query = {}
products = MONGO.mydb.items.find(mongo_query, no_cursor_timeout=True)
if products:
pcount = products.count()
i = 1
print "Procesando %s productos..." % pcount
for p in products:
try:
queue.put((p, i, pcount))
i += 1
except Exception, e:
utils.log(e)
continue
queue.put('STOP')
def main():
queue = multiprocessing.Queue()
procs = [multiprocessing.Process(target=dowork, args=(queue,)) for _ in range(CONFIG_POOL_SIZE)]
for p in procs:
p.start()
populate_jobs(queue)
for p in procs:
p.join()
Also, I've noticed that about every 2500 aprox documents, script pauses for about .5 - 1 secs which is obviously a bad issue. This is a MongoDB problem becase if I do the exactly same loop but using a range(0, 1000000) script doesn't pause at all and runs at 57,000 iterations per second, with a total of 20 seconds to end the script... Huge difference from 2,800 MongoDB documents per second...
This is the code to run a 1,000,000 iteration loop instead docs.
def populate_jobs(queue):
mongo_query = {}
products = MONGO.mydb.items.find(mongo_query, no_cursor_timeout=True)
if products:
pcount = 1000000
i = 1
print "Procesando %s productos..." % pcount
for p in range(0, 1000000):
queue.put((p, i, pcount))
i += 1
queue.put('STOP')
UPDATE
As I saw, the problem is not the multiprocessing itself, is the cursor filling the Queue which is not running in multiprocessing mode, it is one simple process that fills the Queue (populateJobs method) maybe if I could make the cursor multithread/multirpocess and fill the Queue in parallel it will be filled up faster, then the multiprocessing method dowork will do faster, because I think there's a bottleneck where I only fill about 2,800 items per second in Queue and retrieving a lot more in dowork multiprocess, but I don't know how can I parallelize MongoDB cursor.
Maybe, the problem is the latency between my computer and the server's MongoDB. That latency, between me asking for next cursor and MongoDB telling me which is, reduces my performance by 2000% (from 61,000 str/s to 2,800 doc/s)
NOPE I've tried on a localhost MongoDB and performance is exactly the same... This is driving me nuts
Here's how you can use a Pool to feed the children:
START = time.time()
def remaining_time(a, b):
if START:
y = (time.time() - START)
z = ((a * y) / b) - y
d = time.strftime('%H:%M:%S', time.gmtime(z))
e = round(b / y)
progress("{0}/{1} | Tiempo restante {2} ({3}p/s)".format(b, a, d, e), b, a)
def progress(p, c, t):
pc = (c * 100) / t
sys.stdout.write("%s [%-20s] %d%%\r" % (p, '█' * (pc / 5), pc))
sys.stdout.flush()
def dowork(args):
p, i, pcount = args
remaining_time(pcount, i)
def main():
queue = multiprocessing.Queue()
procs = [multiprocessing.Process(target=dowork, args=(queue,)) for _ in range(CONFIG_POOL_SIZE)]
pool = multiprocessing.Pool(CONFIG_POOL_SIZE)
mongo_query = {}
products = MONGO.mydb.items.find(mongo_query, no_cursor_timeout=True)
pcount = products.count()
pool.map(dowork, ((p, idx, pcount) for idx,p in enumerate(products)))
pool.close()
pool.join()
Note that using pool.map requires loading everything from the cursor into memory at once, though, which might be a problem because of how large it is. You can use imap to avoid consuming the whole thing at once, but you'll need to specify a chunksize to minimize IPC overhead:
# Calculate chunksize using same algorithm used internally by pool.map
chunksize, extra = divmod(pcount, CONFIG_POOL_SIZE * 4)
if extra:
chunksize += 1
pool.imap(dowork, ((p, idx, pcount) for idx,p in enumerate(products)), chunksize=chunksize)
pool.close()
pool.join()
For 1,000,000 items, that gives a chunksize of 12,500. You can try sizes larger and smaller than that, and see how it affects performance.
I'm not sure this will help much though, if the bottleneck is actually just pulling the data out of MongoDB.
Why are you using multiprocessing? You don't seem to be doing actual work in other threads using the queue. Python has a global interpreter lock which makes multithreaded code less performant than you'd expect. It's probably making this program slower, not faster.
A couple performance tips:
Try setting batch_size in your find() call to some big number (e.g. 20000). This is the maximum number of documents returned at a time, before the client fetches more, and the default is 101.
Try setting cursor_type to pymongo.cursor.CursorType.EXHAUST, which might reduce the latency you're seeing.