We Keep Coding

Is this the most I can get from Python multiprocess?

I have data, which is in a text file. Each line is a computation to do. This file has around 100 000 000 lines.
First I load everything into the ram, then I have a a method that performs the computation and gives the following results:
def process(data_line):
#do computation
return result
Then I call it like this with packets of 2000 lines and then save the result to disk :
POOL_SIZE = 15 #nbcore - 1
PACKET_SIZE = 2000
pool = Pool(processes=POOL_SIZE)
data_lines = util.load_data_lines(to_be_computed_filename)
number_of_packets = int(number_of_lines/ PACKET_SIZE)
for i in range(number_of_packets):
lines_packet = data_lines[:PACKET_SIZE]
data_lines = data_lines[PACKET_SIZE:]
results = pool.map(process, lines_packet)
save_computed_data_to_disk(to_be_computed_filename, results)
# process the last packet, which is smaller
results.extend(pool.map(process, data_lines))
save_computed_data_to_disk(to_be_computed_filename, results)
print("Done")
The problem is, while I was writing to disk, my CPU is computing nothing and has 8 cores. It is looking at the task manager and it seems that quite a lot of CPU time is lost.
I have to write to disk after having completed my computation because the results are 1000 times larger than the input.
Anyways, I would have to write to the disk at some point. If time is not lost here, it will be lost later.
What could I do to allow one core to write to disk, while still computing with the others? Switch to C?
At this rate I can process 100 millions lines in 75h, but I have 12 billions lines to process, so any improvement is welcome.
example of timings:
Processing packet 2/15 953 of C:/processing/drop_zone\to_be_processed_txt_files\t_to_compute_303620.txt
Launching task and waiting for it to finish...
Task completed, Continuing
Packet was processed in 11.534576654434204 seconds
We are currently going at a rate of 0.002306915330886841 sec/words
Which is 433.47928145051293 words per seconds
Saving in temporary file
Printing writing 5000 computed line to disk took 0.04400920867919922 seconds
saving word to resume from : 06 20 25 00 00
Estimated time for processing the remaining packets is : 51:19:25

Note: This SharedMemory works only for Python >= 3.8 since it first appeared there
Start 3 kinds of processes: Reader, Processor(s), Writer.
Have Reader process read the file incrementally, sharing the result via shared_memory and Queue.
Have the Processor(s) consume the Queue, consume the shared_memory, and return the result(s) via another Queue. Again, as shared_memory.
Have the Writer process consume the second Queue, writing to the destination file.
Have them all communicate through, say, some Events or DictProxy, with the MainProcess who will act as the orchestrator.
Example:
import time
import random
import hashlib
import multiprocessing as MP
from queue import Queue, Empty
# noinspection PyCompatibility
from multiprocessing.shared_memory import SharedMemory
from typing import Dict, List
def readerfunc(
shm_arr: List[SharedMemory], q_out: Queue, procr_ready: Dict[str, bool]
):
numshm = len(shm_arr)
for batch in range(1, 6):
print(f"Reading batch #{batch}")
for shm in shm_arr:
#### Simulated Reading ####
for j in range(0, shm.size):
shm.buf[j] = random.randint(0, 255)
#### ####
q_out.put((batch, shm))
# Need to sync here because we're reusing the same SharedMemory,
# so gotta wait until all processors are done before sending the
# next batch
while not q_out.empty() or not all(procr_ready.values()):
time.sleep(1.0)
def processorfunc(
q_in: Queue, q_out: Queue, suicide: type(MP.Event()), procr_ready: Dict[str, bool]
):
pname = MP.current_process().name
procr_ready[pname] = False
while True:
time.sleep(1.0)
procr_ready[pname] = True
if q_in.empty() and suicide.is_set():
break
try:
batch, shm = q_in.get_nowait()
except Empty:
continue
print(pname, "got batch", batch)
procr_ready[pname] = False
#### Simulated Processing ####
h = hashlib.blake2b(shm.buf, digest_size=4, person=b"processor")
time.sleep(random.uniform(5.0, 7.0))
#### ####
q_out.put((pname, h.hexdigest()))
def writerfunc(q_in: Queue, suicide: type(MP.Event())):
while True:
time.sleep(1.0)
if q_in.empty() and suicide.is_set():
break
try:
pname, digest = q_in.get_nowait()
except Empty:
continue
print("Writing", pname, digest)
#### Simulated Writing ####
time.sleep(random.uniform(3.0, 6.0))
#### ####
print("Writing", pname, digest, "done")
def main():
shm_arr = [
SharedMemory(create=True, size=1024)
for _ in range(0, 5)
]
q_read = MP.Queue()
q_write = MP.Queue()
procr_ready = MP.Manager().dict()
poison = MP.Event()
poison.clear()
reader = MP.Process(target=readerfunc, args=(shm_arr, q_read, procr_ready))
procrs = []
for n in range(0, 3):
p = MP.Process(
target=processorfunc, name=f"Proc{n}", args=(q_read, q_write, poison, procr_ready)
)
procrs.append(p)
writer = MP.Process(target=writerfunc, args=(q_write, poison))
reader.start()
[p.start() for p in procrs]
writer.start()
reader.join()
print("Reader has ended")
while not all(procr_ready.values()):
time.sleep(5.0)
poison.set()
[p.join() for p in procrs]
print("Processors have ended")
writer.join()
print("Writer has ended")
[shm.close() for shm in shm_arr]
[shm.unlink() for shm in shm_arr]
if __name__ == '__main__':
main()

You say you have 8 cores, yet you have:
POOL_SIZE = 15 #nbcore - 1
Assuming you want to leave one processor free (presumably for the main process?) why wouldn't this number be 7? But why do you even want to read a processor free? You are making successive calls to map. While the main process is waiting for these calls to return, it requires know CPU. This is why if you do not specify a pool size when you instantiate your pool it defaults to the number of CPUs you have and not that number minus one. I will have more to say about this below.
Since you have a very large, in-memory list, is it possible that you are expending waisted cycles in your loop rewriting this list on each iteration of the loop. Instead, you can just take a slice of the list and pass that as the iterable argument to map:
POOL_SIZE = 15 # ????
PACKET_SIZE = 2000
data_lines = util.load_data_lines(to_be_computed_filename)
number_of_packets, remainder = divmod(number_of_lines, PACKET_SIZE)
with Pool(processes=POOL_SIZE) as pool:
offset = 0
for i in range(number_of_packets):
results = pool.map(process, data_lines[offset:offset+PACKET_SIZE])
offset += PACKET_SIZE
save_computed_data_to_disk(to_be_computed_filename, results)
if remainder:
results = pool.map(process, data_lines[offset:offset+remainder])
save_computed_data_to_disk(to_be_computed_filename, results)
print("Done")
Between each call to map the main process is writing out the results to to_be_computed_filename. In the meanwhile, every process in your pool is sitting idle. This should be given to another process (actually a thread running under the main process):
import multiprocessing
import queue
import threading
POOL_SIZE = 15 # ????
PACKET_SIZE = 2000
data_lines = util.load_data_lines(to_be_computed_filename)
number_of_packets, remainder = divmod(number_of_lines, PACKET_SIZE)
def save_data(q):
while True:
results = q.get()
if results is None:
return # signal to terminate
save_computed_data_to_disk(to_be_computed_filename, results)
q = queue.Queue()
t = threading.Thread(target=save_data, args=(q,))
t.start()
with Pool(processes=POOL_SIZE) as pool:
offset = 0
for i in range(number_of_packets):
results = pool.map(process, data_lines[offset:offset+PACKET_SIZE])
offset += PACKET_SIZE
q.put(results)
if remainder:
results = pool.map(process, data_lines[offset:offset+remainder])
q.put(results)
q.put(None)
t.join() # wait for thread to terminate
print("Done")
I've chosen to run save_data in a thread of the main process. This could also be another process in which case you would need to use a multiprocessing.Queue instance. But I figured the main process thread is mostly waiting for the map to complete and there would not be competition for the GIL. Now if you do not leave a processor free for the threading job, save_data, it may end up doing most of the saving only after all of the results have been created. You would need to experiment a bit with this.
Ideally, I would also modify the reading of the input file so as to not have to first read it all into memory but rather read it line by line yielding 2000 line chunks and submitting those as jobs for map to process:
import multiprocessing
import queue
import threading
POOL_SIZE = 15 # ????
PACKET_SIZE = 2000
def save_data(q):
while True:
results = q.get()
if results is None:
return # signal to terminate
save_computed_data_to_disk(to_be_computed_filename, results)
def read_data():
"""
yield lists of PACKET_SIZE
"""
lines = []
with open(some_file, 'r') as f:
for line in iter(f.readline(), ''):
lines.append(line)
if len(lines) == PACKET_SIZE:
yield lines
lines = []
if lines:
yield lines
q = queue.Queue()
t = threading.Thread(target=save_data, args=(q,))
t.start()
with Pool(processes=POOL_SIZE) as pool:
for l in read_data():
results = pool.map(process, l)
q.put(results)
q.put(None)
t.join() # wait for thread to terminate
print("Done")

I made two assumptions: The writing is hitting the I/O bound, not the CPU bound - meaning that throwing more cores onto writing would not improve the performance. And the process function contains some heavy computations.
I would approach it differently:
Split up the large list into a list of list
Feed it than into the processes
Store the total result
Here is the example code:
import multiprocessing as mp
data_lines = [0]*10000 # read it from file
size = 2000
# Split the list into a list of list (with chunksize `size`)
work = [data_lines[i:i + size] for i in range(0, len(data_lines), size)]
def process(data):
result = len(data) # some something fancy
return result
with mp.Pool() as p:
result = p.map(process, work)
save_computed_data_to_disk(file_name, result)
On meta: You may also have a look into numpy or pandas (depending on the data) because it sounds that you would like to do something into that direction.

The first thing that comes to mind for the code is to run the saving function in the thread. By this we exclude the bottelneck of waiting disk writing. Like so:
executor = ThreadPoolExecutor(max_workers=2)
future = executor.submit(save_computed_data_to_disk, to_be_computed_filename, results)
saving_futures.append(future)
...
concurrent.futures.wait(saving_futures, return_when=ALL_COMPLETED) # wait all saved to disk after processing
print("Done")

Producer-consumer problem - trying to save into a csv file

so this seemingly simple problem is doing my head in.
I have a dataset (datas) and I do some processing on it (this isn't the issue, though this takes time owing to the size of the dataset) to produce multiple rows to be stored into a CSV file. However, it is very taxing to produce a row, then save it to csv, then produce a row and then save it etc.
So I'm trying to implement producer and consumer threads - producers will produce each row of data (to speed up the process), store in a queue and a single consumer will then append to my csv file.
My attempts below result in success sometimes (the data is correctly saved) or other times the data is "cut off" (either an entire row or part of it).
What am I doing wrong?
from threading import Thread
from queue import Queue
import csv
q = Queue()
def producer():
datas = [["hello","world"],["test","hey"],["my","away"],["your","gone"],["bye","hat"]]
for data in datas:
q.put(data)
def consumer():
while True:
local = q.get()
file = open('dataset.csv','a')
with file as fd:
writer = csv.writer(fd)
writer.writerow(local)
file.close()
q.task_done()
for i in range(10):
t = Thread(target=consumer)
t.daemon = True
t.start()
producer()
q.join()

I think this does something similar to what you're trying to do. For testing purposes, it prefixes each row of data in the CSV file produced with a "producer id" so the source of the data can be seen in the results.
As you will be able to see from the csv file produced, all the data produced gets put into it.
import csv
import random
from queue import Queue
from threading import Thread
import time
SENTINEL = object()
def producer(q, id):
data = (("hello", "world"), ("test", "hey"), ("my", "away"), ("your", "gone"),
("bye", "hat"))
for datum in data:
q.put((id,) + datum) # Prefix producer ID to datum for testing.
time.sleep(random.random()) # Vary thread speed for testing.
class Consumer(Thread):
def __init__(self, q):
super().__init__()
self.q = q
def run(self):
with open('dataset.csv', 'w', newline='') as file:
writer = csv.writer(file, delimiter=',')
while True:
datum = self.q.get()
if datum is SENTINEL:
break
writer.writerow(datum)
def main():
NUM_PRODUCERS = 10
queue = Queue()
# Create producer threads.
threads = []
for id in range(NUM_PRODUCERS):
t = Thread(target=producer, args=(queue, id+1,))
t.start()
threads.append(t)
# Create Consumer thread.
consumer = Consumer(queue)
consumer.start()
# Wait for all producer threads to finish.
while threads:
threads = [thread for thread in threads if thread.is_alive()]
queue.put(SENTINEL) # Indicate to consumer thread no more data.
consumer.join()
print('Done')
if __name__ == '__main__':
main()

How do I restart a thread if its duration is longer than 10 seconds?

How in the main thread can I track the duration of the function write_file()?
Task: create a condition, if the execution time of the function is more than 10 seconds, then it is necessary to restart the function.
from multiprocessing import Pool
def write_file(file: str):
f = open(file, 'w')
for item in range(0, 1500000):
f.write("%s\n" % item)
f.close()
if __name__ == '__main__':
list_files = ['1.txt', '2.txt', '3.txt']
with Pool(3) as p:
p.map(write_file, list_files)

I found attempt to amend Pool to be overcomplicated here.
The Pool class let workers to be alive untill the whole working queue is done and thus has complex mechanism to control it.
Instead, if you have not very stict requirenments on 10 second, you can use the following code:
from multiprocessing import Process
import time
pdict = {}
for fname in list_files:
p = Process(target = write_file, args = (fname,))
pdict[fname] = p
p.start()
while pdict:
to_del = []
time.sleep(10)
for pname in pdict:
if pdict[pname].exitcode == None or pdict[pname].is_alive():
pdict[pname].terminate() #killing old; that should also release file resource
pdict[pname] = Process(target = write_file, args = (pname,))
pdict[pname].start() #simply creating new and starting
else:
to_del.append(pname)
for pname in to_del:
del pdict[pname]

How can I terminate running jobs without closing connection to the core? (currently using execnet)

I have a cluster of computers which uses a master node to communicate with the slave nodes in the cluster.
The main problem I'm facing is using execnet is being able to kill certain jobs that are running and then having new jobs requeue on the same core that the other job just got terminated on (as I want to utilize all cores of the slave nodes at any given time).
As of now there is no way to terminate running jobs using execnet, so I figured if I could just kill the jobs manually through a bash script, say sudo kill 12345 where 12345 is the PID of the job (obtaining the PID of each job is another thing not supported by execnet, but that's another topic), then it would terminate the job and then requeue another on the same core that was just terminated on. It does kill the job correctly, however it closes the connection to that channel (the core; the master node communicates to each core individually) and then does not utilize that core anymore, until all jobs are done. Is there a way to terminate a running job, without killing the connection to the core?
Here is the script to submit jobs
import execnet, os, sys
import re
import socket
import numpy as np
import pickle, cPickle
from copy import deepcopy
import time
import job
def main():
print 'execnet source files are located at:\n {}/\n'.format(
os.path.join(os.path.dirname(execnet.__file__))
)
# Generate a group of gateways.
work_dir = '/home/mpiuser/pn2/'
f = 'cluster_core_info.txt'
n_start, n_end = 250000, 250008
ci = get_cluster_info(f)
group, g_labels = make_gateway_group(ci, work_dir)
mch = group.remote_exec(job)
args = range(n_start, n_end+1) # List of parameters to compute factorial.
manage_jobs(group, mch, queue, g_labels, args)
# Close the group of gateways.
group.terminate()
def get_cluster_info(f):
nodes, ncores = [], []
with open(f, 'r') as fid:
while True:
line = fid.readline()
if not line:
fid.close()
break
line = line.strip('\n').split()
nodes.append(line[0])
ncores.append(int(line[1]))
return dict( zip(nodes, ncores) )
def make_gateway_group(cluster_info, work_dir):
''' Generate gateways on all cores in remote nodes. '''
print 'Gateways generated:\n'
group = execnet.Group()
g_labels = []
nodes = list(cluster_info.keys())
for node in nodes:
for i in range(cluster_info[node]):
group.makegateway(
"ssh={0}//id={0}_{1}//chdir={2}".format(
node, i, work_dir
))
sys.stdout.write(' ')
sys.stdout.flush()
print list(group)[-1]
# Generate a string 'node-id_core-id'.
g_labels.append('{}_{}'.format(re.findall(r'\d+',node)[0], i))
print ''
return group, g_labels
def get_mch_id(g_labels, string):
ids = [x for x in re.findall(r'\d+', string)]
ids = '{}_{}'.format(*ids)
return g_labels.index(ids)
def manage_jobs(group, mch, queue, g_labels, args):
args_ref = deepcopy(args)
terminated_channels = 0
active_jobs, active_args = [], []
while True:
channel, item = queue.get()
if item == 'terminate_channel':
terminated_channels += 1
print " Gateway closed: {}".format(channel.gateway.id)
if terminated_channels == len(mch):
print "\nAll jobs done.\n"
break
continue
if item != "ready":
mch_id_completed = get_mch_id(g_labels, channel.gateway.id)
depopulate_list(active_jobs, mch_id_completed, active_args)
print " Gateway {} channel id {} returned:".format(
channel.gateway.id, mch_id_completed)
print " {}".format(item)
if not args:
print "\nNo more jobs to submit, sending termination request...\n"
mch.send_each(None)
args = 'terminate_channel'
if args and \
args != 'terminate_channel':
arg = args.pop(0)
idx = args_ref.index(arg)
channel.send(arg) # arg is copied by value to the remote side of
# channel to be executed. Maybe blocked if the
# sender queue is full.
# Get the id of current channel used to submit a job,
# this id can be used to refer mch[id] to terminate a job later.
mch_id_active = get_mch_id(g_labels, channel.gateway.id)
print "Job {}: {}! submitted to gateway {}, channel id {}".format(
idx, arg, channel.gateway.id, mch_id_active)
populate_list(active_jobs, mch_id_active,
active_args, arg)
def populate_list(jobs, job_active, args, arg_active):
jobs.append(job_active)
args.append(arg_active)
def depopulate_list(jobs, job_completed, args):
i = jobs.index(job_completed)
jobs.pop(i)
args.pop(i)
if __name__ == '__main__':
main()
and here is my job.py script:
#!/usr/bin/env python
import os, sys
import socket
import time
import numpy as np
import pickle, cPickle
import random
import job
def hostname():
return socket.gethostname()
def working_dir():
return os.getcwd()
def listdir(path):
return os.listdir(path)
def fac(arg):
return np.math.factorial(arg)
def dump(arg):
path = working_dir() + '/out'
if not os.path.exists(path):
os.mkdir(path)
f_path = path + '/fac_{}.txt'.format(arg)
t_0 = time.time()
num = fac(arg) # Main operation
t_1 = time.time()
cPickle.dump(num, open(f_path, "w"), protocol=2) # Main operation
t_2 = time.time()
duration_0 = "{:.4f}".format(t_1 - t_0)
duration_1 = "{:.4f}".format(t_2 - t_1)
#num2 = cPickle.load(open(f_path, "rb"))
return '--Calculation: {} s, dumping: {} s'.format(
duration_0, duration_1)
if __name__ == '__channelexec__':
channel.send("ready")
for arg in channel:
if arg is None:
break
elif str(arg).isdigit():
channel.send((
str(arg)+'!',
job.hostname(),
job.dump(arg)
))
else:
print 'Warnning! arg sent should be number | None'

Yes, you are on the right track. Use psutil library to manage the processes, find their pids etc.
And kill them. No need for involveing bash anywhere. Python covers it all.
Or, even better, program your script to terminate when master say so.
It is usually done that way.
You can even make it start another script before terminating itself if you want/need.
Or, if it is the same that you would be doing in another process, just stop the current work and start a new one in the script without terminating it at all.
And, if I may make a suggestion. Don't read your file line by line, read a whole file and then use *.splitlines(). For small files reading them in chunks just tortures the IO. You wouldn't be needing *.strip() as well. And you should remove unused imports too.

Multiprocessing, writing to file, and deadlock on large loops

I have a very weird problem with the code below. when numrows = 10 the Process loops completes itself and proceeds to finish. If the growing list becomes larger it goes into a deadlock. Why is this and how can I solve this?
import multiprocessing, time, sys
# ----------------- Calculation Engine -------------------
def feed(queue, parlist):
for par in parlist:
queue.put(par)
def calc(queueIn, queueOut):
while True:
try:
par = queueIn.get(block = False)
print "Project ID: %s started. " % par
res = doCalculation(par)
queueOut.put(res)
except:
break
def write(queue, fname):
print 'Started to write to file'
fhandle = open(fname, "w")
while True:
try:
res = queue.get(block = False)
for m in res:
print >>fhandle, m
except:
break
fhandle.close()
print 'Complete writing to the file'
def doCalculation(project_ID):
numrows = 100
toFileRowList = []
for i in range(numrows):
toFileRowList.append([project_ID]*100)
print "%s %s" % (multiprocessing.current_process().name, i)
return toFileRowList
def main():
parlist = [276, 266]
nthreads = multiprocessing.cpu_count()
workerQueue = multiprocessing.Queue()
writerQueue = multiprocessing.Queue()
feedProc = multiprocessing.Process(target = feed , args = (workerQueue, parlist))
calcProc = [multiprocessing.Process(target = calc , args = (workerQueue, writerQueue)) for i in range(nthreads)]
writProc = multiprocessing.Process(target = write, args = (writerQueue, 'somefile.csv'))
feedProc.start()
feedProc.join ()
for p in calcProc:
p.start()
for p in calcProc:
p.join()
writProc.start()
writProc.join()
if __name__=='__main__':
sys.exit(main())

I think the problem is the Queue buffer getting filled, so you need to read from the queue before you can put additional stuff in it.
For example, in your feed thread you have:
queue.put(par)
If you keep putting much stuff without reading this will cause it to block untill the buffer is freed, but the problem is that you only free the buffer in your calc thread, which in turn doesn't get started before you join your blocking feed thread.
So, in order for your feed thread to finish, the buffer should be freed, but the buffer won't be freed before the thread finishes :)
Try organizing your queues access more.

The feedProc and the writeProc are not actually running in parallel with the rest of your program. When you have
proc.start()
proc.join ()
you start the process and then, on the join() you immediatly wait for it to finish. In this case there's no gain in multiprocessing, only overhead. Try to start ALL processes at once before you join them. This will also have the effect that your queues get emptied regularyl and you won't deadlock.