Develop Reference

Python multiprocessing - sharing large dataset

I'm trying to speed up a CPU-bound Python script (on Windows11). Threats in Python do not seem to run on a different cpu(core) so the only option I have is multiprocessing.
I have a big dictionary data structure (11GB memory footprint after loading from file) that I am checking calculated values on if they are in that dictionary. Input for the calculation also comes from a file (100GB in size). This input I can pool-map to the processes in batches, no problem. But I cannot copy the dictionary to all processes because there is not enough memory for that. So I need to find a way for the processes to check if the value (actually a string) is in the dictionary.
Any advice?
Pseudo programm flow:
--main--
- load dictionary structure from file # 11GB memory footprint
- ...
- While not all chuncks loaded
- Load chunk of calcdata from file # (10.000 lines per chunk)
- Distribute (map) calcdata-chunck to processes
- Wait for processes to complete all chunks
--process--
- for each element in subchunk
- perform calculation
- check if calculation in dictionary # here is my problem!
- store result in file
Edit, after implementing comments below, I am now at:
def ReadDictFromFile()
cnt=0
print("Reading dictionary from " + dictfilename)
with open(dictfilename, encoding=("utf-8"), errors=("replace")) as f:
next(f) #skip first line (header)
for line in f:
s = line.rstrip("\n")
(key,keyvalue) = s.split()
shared_dict[str(key)]=keyvalue
cnt = cnt + 1
if ((cnt % 1000000) == 0): #log each 1000000 where we are
print(cnt)
return #temp to speed up testing, not load whole dictionary atm
print("Done loading dictionary")
def checkqlist(qlist)
print(str(os.getpid()) + "-" + str(len(qlist)))
for li in qlist:
try:
checkvalue = calculations(li)
(found, keyval) = InMem(checkvalue)
if (found):
print("FOUND!!! " + checkvalue + ' ' + keyvalue)
except Exception as e:
print("(" + str(os.getpid()) + ")Error log: %s" % repr(e))
time.sleep(15)
def InMem(checkvalue):
if(checkvalue in shared_dict):
return True, shared_dict[checkvalue]
else:
return False, ""
if __name__ == "__main__":
start_time = time.time()
global shared_dict
manager = Manager()
shared_dict = manager.dict()
ReadDictFromFile()
chunksize=5
nr_of_processes = 10
with open(filetocheck, encoding=("utf-8"), errors=("replace")) as f:
qlist = []
for line in f:
s = line.rstrip("\n")
qlist.append(s)
if (len(qlist) >= (chunksize * nr_of_processes)):
chunked_list = [qlist[i:i+chunk_size] for i in range(0, len(qlist), chunk_size)]
try:
with multiprocessing.Pool() as pool:
pool.map(checkqlist, chunked_list, nr_of_processes) #problem: qlist is a single string, not a list of about 416 strings.
except Exception as e:
print("error log: %s" % repr(e))
time.sleep(15)
logit("Completed! " + datetime.datetime.now().strftime("%I:%M%p on %B %d, %Y"))
print("--- %s seconds ---" % (time.time() - start_time))

you can use a multiprocessing.Manager.dict for this, it's the fastest IPC you can use to do the check between processes in python, and for the memory size, just make it smaller by changing all values to None, on my pc it can do 33k member checks every second ... about 400 times slower than a normal dictionary.
manager = Manager()
shared_dict = manager.dict()
shared_dict.update({x:None for x in main_dictionary})
shared_dict["new_element"] = None # to set another value
del shared_dict["new_element"] # to delete a certain value
you can also use a dedicated in-memory database for this like redis, which can handle being polled by multiple processes at the same time.
#Sam Mason suggestion to use WSL and fork may be better, but this one is the most portable.
Edit: to store it in children global scope you have to pass it through the initializer.
def define_global(var):
global shared_dict
shared_dict = var
...
if __name__ == "__main__":
...
with multiprocessing.Pool(initializer=define_global, initargs=(shared_dict ,)) as pool:

Python end function after another function returns result

I may be approaching this all wrong but still this is where I'm at. I have very large log files I'm trying to search, up to 30gb in some cases. I'm writing a script to pull info and have been playing with multi process to speed it up a bit. right now I'm testing running two functions at the same time to search from the top and bottom to get results, which seems to work. I'm wondering if it's possible to stop one function one a result from the other. Such as if the top function finds a result they both stop. This way I can build it out as needed.
from file_read_backwards import FileReadBackwards
from multiprocessing import Process
import sys
z = "log.log"
#!/usr/bin/env python
rocket = 0
def top():
target = "test"
with open(z) as src:
found= None
for line in src:
if len(line) == 0: break #happens at end of file, then stop loop
if target in line:
found= line
break
print(found)
def bottom():
target = "text"
with FileReadBackwards(z) as src:
found= None
for line in src:
if len(line) == 0: break #happens at end of file, then stop loop
if target in line:
found= line
break
print(found)
if __name__=='__main__':
p1 = Process(target = top)
p1.start()
p2 = Process(target = bottom)
p2.start()

Here's a proof-of-concept of the approach I mentioned in the comments:
import os
import random
import sys
from multiprocessing import Process, Value
def search(proc_no, file_name, seek_to, max_size, find, flag):
stop_at = seek_to + max_size
with open(file_name) as f:
if seek_to:
f.seek(seek_to - 1)
prev_char = f.read(1)
if prev_char != '\n':
# Landed in the middle of a line. Skip back one (or
# maybe more) lines so this line isn't excluded. Start
# by seeking back 256 bytes, then 512 if necessary, etc.
exponent = 8
pos = seek_to
while pos >= seek_to:
pos = f.seek(max(0, pos - (2 ** exponent)))
f.readline()
pos = f.tell()
exponent += 1
while True:
if flag.value:
break
line = f.readline()
if not line:
break # EOF
data = line.strip()
if data == find:
flag.value = proc_no
print(data)
break
if f.tell() > stop_at:
break
if __name__ == '__main__':
# list.txt contains lines with the numbers 1 to 1000001
file_name = 'list.txt'
info = os.stat(file_name)
file_size = info.st_size
if len(sys.argv) == 1:
# Pick a random value from list.txt
num_lines = 1000001
choices = list(range(1, num_lines + 1))
choices.append('XXX')
find = str(random.choice(choices))
else:
find = sys.argv[1]
num_procs = 4
chunk_size, remainder = divmod(file_size, num_procs)
max_size = chunk_size + remainder
flag = Value('i', 0)
procs = []
print(f'Using {num_procs} processes to look for {find} in {file_name}')
for i in range(num_procs):
seek_to = i * chunk_size
proc = Process(target=search, args=(i + 1, file_name, seek_to, max_size, find, flag))
procs.append(proc)
for proc in procs:
proc.start()
for proc in procs:
proc.join()
if flag.value:
print(find, 'found by proc', flag.value)
else:
print(find, 'not found')
After reading various posts[1] about reading files with multiprocessing and multithreading, it seems that neither is a great approach due to potential disk thrashing and serialized reads. So here's a different, simpler approach that is way faster (at least for the file with a million lines I was trying it out on):
import mmap
import sys
def search_file(file_name, text, encoding='utf-8'):
text = text.encode(encoding)
with open(file_name) as f:
with mmap.mmap(f.fileno(), 0, flags=mmap.ACCESS_READ, prot=mmap.PROT_READ) as m:
index = m.find(text)
if index > -1:
# Found a match; now find beginning of line that
# contains match so we can grab the whole line.
while index > 0:
index -= 1
if m[index] == 10:
index += 1
break
else:
index = 0
m.seek(index)
line = m.readline()
return line.decode(encoding)
if __name__ == '__main__':
file_name, search_string = sys.argv[1:]
line = search_file(file_name, search_string)
sys.stdout.write(line if line is not None else f'Not found in {file_name}: {search_string}\n')
I'm curious how this would perform with a 30GB log file.
[1] Including this one

Simple example using a multiprocessing.Pool and callback function.
Terminates remaining pool processes once a result has returned.
You could add an arbitrary number of processes to search from different offsets in the file using this approach.
import math
import time
from multiprocessing import Pool
from random import random
def search(pid, wait):
"""Sleep for wait seconds, return PID
"""
time.sleep(wait)
return pid
def done(result):
"""Do something with result and stop other processes
"""
print("Process: %d done." % result)
pool.terminate()
print("Terminate Pool")
pool = Pool(2)
pool.apply_async(search, (1, math.ceil(random() * 3)), callback=done)
pool.apply_async(search, (2, math.ceil(random() * 3)), callback=done)
# do other stuff ...
# Wait for result
pool.close()
pool.join() # block our main thread

This is essentially the same as Blurp's answer, but I shortened it and made it a bit to make it more general. As you can see top should be an infinite loop, but bottom stops top immediately.
from multiprocessing import Process
valNotFound = True
def top():
i=0
while ValNotFound:
i += 1
def bottom():
ValNotFound = False
p1 = Process(target = top)
p2 = Process(target = bottom)
p1.start()
p2.start()

Pytonic way of passing values between process

I need a simple way to pass the stdout of a subprocess as a list to another function using multiprocess:
The first function that invokes subprocess:
def beginRecvTest():
command = ["receivetest","-f=/dev/pcan33"]
incoming = Popen(command, stdout = PIPE)
processing = iter(incoming.stdout.readline, "")
lines = list(processing)
return lines
The function that should receive lines:
def readByLine(lines):
i = 0
while (i < len(lines)):
system("clear")
if(lines[i][0].isdigit()):
line = lines[i].split()
dictAdd(line)
else:
next
print ; print "-" *80
for _i in mydict.keys():
printMsg(mydict, _i)
print "Keys: ", ; print mydict.keys()
print ; print "-" *80
sleep(0.3)
i += 1
and the main from my program:
if __name__ == "__main__":
dataStream = beginRecvTest()
p = Process(target=dataStream)
reader = Process(target=readByLine, args=(dataStream,))
p.start()
reader.start()
I've read up on using queues, but I don't think that's exactly what I need.
The subprocess called returns infinite data so some people have suggested using tempfile, but I am totally confused about how to do this.
At the moment the script only returns the first line read, and all efforts on looping the beginRecvTest() function have ended in compilation errors.

Successive multiprocessing

I am filtering huge text files using multiprocessing.py. The code basically opens the text files, works on it, then closes it.
Thing is, I'd like to be able to launch it successively on multiple text files. Hence, I tried to add a loop, but for some reason it doesn't work (while the code works on each file). I believe this is an issue with:
if __name__ == '__main__':
However, I am looking for something else. I tried to create a Launcher and a LauncherCount files like this:
LauncherCount.py:
def setLauncherCount(n):
global LauncherCount
LauncherCount = n
and,
Launcher.py:
import os
import LauncherCount
LauncherCount.setLauncherCount(0)
os.system("OrientedFilterNoLoop.py")
LauncherCount.setLauncherCount(1)
os.system("OrientedFilterNoLoop.py")
...
I import LauncherCount.py, and use LauncherCount.LauncherCount as my loop index.
Of course, this doesn't work too as it edits the variable LauncherCount.LauncherCount locally, so it won't be edited in the imported version of LauncherCount.
Is there any way to edit globally a variable in an imported file? Or, is there any way to do this in any other way? What I need is running a code multiple times, in changing one value, and without using any loop apparently.
Thanks!
Edit: Here is my main code if necessary. Sorry for the bad style ...
import multiprocessing
import config
import time
import LauncherCount
class Filter:
""" Filtering methods """
def __init__(self):
print("launching methods")
# Return the list: [Latitude,Longitude] (elements are floating point numbers)
def LatLong(self,line):
comaCount = []
comaCount.append(line.find(','))
comaCount.append(line.find(',',comaCount[0] + 1))
comaCount.append(line.find(',',comaCount[1] + 1))
Lat = line[comaCount[0] + 1 : comaCount[1]]
Long = line[comaCount[1] + 1 : comaCount[2]]
try:
return [float(Lat) , float(Long)]
except ValueError:
return [0,0]
# Return a boolean:
# - True if the Lat/Long is within the Lat/Long rectangle defined by:
# tupleFilter = (minLat,maxLat,minLong,maxLong)
# - False if not
def LatLongFilter(self,LatLongList , tupleFilter) :
if tupleFilter[0] <= LatLongList[0] <= tupleFilter[1] and
tupleFilter[2] <= LatLongList[1] <= tupleFilter[3]:
return True
else:
return False
def writeLine(self,key,line):
filterDico[key][1].write(line)
def filteringProcess(dico):
myFilter = Filter()
while True:
try:
currentLine = readFile.readline()
except ValueError:
break
if len(currentLine) ==0: # Breaks at the end of the file
break
if len(currentLine) < 35: # Deletes wrong lines (too short)
continue
LatLongList = myFilter.LatLong(currentLine)
for key in dico:
if myFilter.LatLongFilter(LatLongList,dico[key][0]):
myFilter.writeLine(key,currentLine)
###########################################################################
# Main
###########################################################################
# Open read files:
readFile = open(config.readFileList[LauncherCount.LauncherCount][1], 'r')
# Generate writing files:
pathDico = {}
filterDico = config.filterDico
# Create outputs
for key in filterDico:
output_Name = config.readFileList[LauncherCount.LauncherCount][0][:-4]
+ '_' + key +'.log'
pathDico[output_Name] = config.writingFolder + output_Name
filterDico[key] = [filterDico[key],open(pathDico[output_Name],'w')]
p = []
CPUCount = multiprocessing.cpu_count()
CPURange = range(CPUCount)
startingTime = time.localtime()
if __name__ == '__main__':
### Create and start processes:
for i in CPURange:
p.append(multiprocessing.Process(target = filteringProcess ,
args = (filterDico,)))
p[i].start()
### Kill processes:
while True:
if [p[i].is_alive() for i in CPURange] == [False for i in CPURange]:
readFile.close()
for key in config.filterDico:
config.filterDico[key][1].close()
print(key,"is Done!")
endTime = time.localtime()
break
print("Process started at:",startingTime)
print("And ended at:",endTime)

To process groups of files in sequence while working on files within a group in parallel:
#!/usr/bin/env python
from multiprocessing import Pool
def work_on(args):
"""Process a single file."""
i, filename = args
print("working on %s" % (filename,))
return i
def files():
"""Generate input filenames to work on."""
#NOTE: you could read the file list from a file, get it using glob.glob, etc
yield "inputfile1"
yield "inputfile2"
def process_files(pool, filenames):
"""Process filenames using pool of processes.
Wait for results.
"""
for result in pool.imap_unordered(work_on, enumerate(filenames)):
#NOTE: in general the files won't be processed in the original order
print(result)
def main():
p = Pool()
# to do "successive" multiprocessing
for filenames in [files(), ['other', 'bunch', 'of', 'files']]:
process_files(p, filenames)
if __name__=="__main__":
main()
Each process_file() is called in sequence after the previous one has been complete i.e., the files from different calls to process_files() are not processed in parallel.

Unpickling mid-stream (python)

I am writing scripts to process (very large) files by repeatedly unpickling objects until EOF. I would like to partition the file and have separate processes (in the cloud) unpickle and process separate parts.
However my partitioner is not intelligent, it does not know about the boundaries between pickled objects in the file (since those boundaries depend on the object types being pickled, etc.).
Is there a way to scan a file for a "start pickled object" sentinel? The naive way would be to attempt unpickling at successive byte offsets until an object is successfully pickled, but that yields unexpected errors. It seems that for certain combinations of input, the unpickler falls out of sync and returns nothing for the rest of the file (see code below).
import cPickle
import os
def stream_unpickle(file_obj):
while True:
start_pos = file_obj.tell()
try:
yield cPickle.load(file_obj)
except (EOFError, KeyboardInterrupt):
break
except (cPickle.UnpicklingError, ValueError, KeyError, TypeError, ImportError):
file_obj.seek(start_pos+1, os.SEEK_SET)
if __name__ == '__main__':
import random
from StringIO import StringIO
# create some data
sio = StringIO()
[cPickle.dump(random.random(), sio, cPickle.HIGHEST_PROTOCOL) for _ in xrange(1000)]
sio.flush()
# read from subsequent offsets and find discontinuous jumps in object count
size = sio.tell()
last_count = None
for step in xrange(size):
sio.seek(step, os.SEEK_SET)
count = sum(1 for _ in stream_unpickle(file_obj))
if last_count is None or count == last_count - 1:
last_count = count
elif count != last_count:
# if successful, these should never print (but they do...)
print '%d elements read from byte %d' % (count, step)
print '(%d elements read from byte %d)' % (last_count, step-1)
last_count = count

The pickletools module has a dis function that shows the opcodes. It shows that there is a STOP opcode that you may be scan for:
>>> import pickle, pickletools, StringIO
>>> s = StringIO.StringIO()
>>> pickle.dump('abc', s)
>>> p = s.getvalue()
>>> pickletools.dis(p)
0: S STRING 'abc'
7: p PUT 0
10: . STOP
highest protocol among opcodes = 0
Note, using the STOP opcode is a bit tricky because the codes are of variable length, but it may serve as a useful hint about where the cutoffs are.
If you control the pickling step on the other end, then you can improve the situation by adding your own unambiguous alternative separator:
>>> sep = '\xDE\xAD\xBE\xEF'
>>> s = StringIO.StringIO()
>>> pickle.dump('abc', s)
>>> s.write(sep)
>>> pickle.dump([10, 20], s)
>>> s.write(sep)
>>> pickle.dump('def', s)
>>> s.write(sep)
>>> pickle.dump([30, 40], s)
>>> p = s.getvalue()
Before unpacking, split into separate pickles using the known separator:
>>> for pick in p.split(sep):
print pickle.loads(pick)
abc
[10, 20]
def
[30, 40]

In the pickled file, some opcodes have an argument -- a data value that follows the opcode. The data values vary in length, and can contain bytes identical to opcodes. Therefore, if you start reading the file from an arbitrary position, you have no way of knowing if you are looking at an opcode or in the middle of an argument. You must read the file from beginning and parse the opcodes.
I cooked up this function that skips one pickle from a file, i.e. reads it and parses opcodes, but does not construct the objects. It seems slightly faster than cPickle.loads on some files I have. You could rewrite this in C for more speed. (after testing this properly)
Then, you can make one pass over the whole file to get the seek position of each pickle.
from pickletools import code2op, UP_TO_NEWLINE, TAKEN_FROM_ARGUMENT1, TAKEN_FROM_ARGUMENT4
from marshal import loads as mloads
def skip_pickle(f):
"""Skip one pickle from file.
'f' is a file-like object containing the pickle.
"""
while True:
code = f.read(1)
if not code:
raise EOFError
opcode = code2op[code]
if opcode.arg is not None:
n = opcode.arg.n
if n > 0:
f.read(n)
elif n == UP_TO_NEWLINE:
f.readline()
elif n == TAKEN_FROM_ARGUMENT1:
n = ord(f.read(1))
f.read(n)
elif n == TAKEN_FROM_ARGUMENT4:
n = mloads('i' + f.read(4))
f.read(n)
if code == '.':
break

Sorry to answer my own question, and thanks to #RaymondHettinger for the idea of adding sentinels.
Here's what worked for me. I created readers and writers that use a sentinel '#S' followed by a data block length at the beginning of each 'record'. The writer has to take care to find any occurrences of '#' in the data being written and double them (into '##'). The reader then uses a look-behind regex to find sentinels, distinct from any matching values that might be in the original stream, and also verify the number of bytes between this sentinel and the subsequent one.
RecordWriter is a context manager (so multiple calls to write() can be encapsulated into a single record if needed). RecordReader is a generator.
Not sure how this is on performance. Any faster/elegant-er solutions are welcome.
import re
import cPickle
from functools import partial
from cStringIO import StringIO
SENTINEL = '#S'
# when scanning look for #S, but NOT ##S
sentinel_pattern = '(?<!#)#S' # uses negative look-behind
sentinel_re = re.compile(sentinel_pattern)
find_sentinel = sentinel_re.search
# when writing replace single # with double ##
write_pattern = '#'
write_re = re.compile(write_pattern)
fix_write = partial(write_re.sub, '##')
# when reading, replace double ## with single #
read_pattern = '##'
read_re = re.compile(read_pattern)
fix_read = partial(read_re.sub, '#')
class RecordWriter(object):
def __init__(self, stream):
self._stream = stream
self._write_buffer = None
def __enter__(self):
self._write_buffer = StringIO()
return self
def __exit__(self, et, ex, tb):
if self._write_buffer.tell():
self._stream.write(SENTINEL) # start
cPickle.dump(self._write_buffer.tell(), self._stream, cPickle.HIGHEST_PROTOCOL) # byte length of user's original data
self._stream.write(fix_write(self._write_buffer.getvalue()))
self._write_buffer = None
return False
def write(self, data):
if not self._write_buffer:
raise ValueError("Must use StreamWriter as a context manager")
self._write_buffer.write(data)
class BadBlock(Exception): pass
def verify_length(block):
fobj = StringIO(block)
try:
stated_length = cPickle.load(fobj)
except (ValueError, IndexError, cPickle.UnpicklingError):
raise BadBlock
data = fobj.read()
if len(data) != stated_length:
raise BadBlock
return data
def RecordReader(stream):
' Read one record '
accum = StringIO()
seen_sentinel = False
data = ''
while True:
m = find_sentinel(data)
if not m:
if seen_sentinel:
accum.write(data)
data = stream.read(80)
if not data:
if accum.tell():
try: yield verify_length(fix_read(accum.getvalue()))
except BadBlock: pass
return
else:
if seen_sentinel:
accum.write(data[:m.start()])
try: yield verify_length(fix_read(accum.getvalue()))
except BadBlock: pass
accum = StringIO()
else:
seen_sentinel = True
data = data[m.end():] # toss
if __name__ == '__main__':
import random
stream = StringIO()
data = [str(random.random()) for _ in xrange(3)]
# test with a string containing sentinel and almost-sentinel
data.append('abc12#jeoht38#SoSooihetS#')
count = len(data)
for i in data:
with RecordWriter(stream) as r:
r.write(i)
size = stream.tell()
start_pos = random.random() * size
stream.seek(start_pos, os.SEEK_SET)
read_data = [s for s in RecordReader(stream)]
print 'Original data: ', data
print 'After seeking to %d, RecordReader returned: %s' % (start_pos, read_data)