Skip to content
/ mplite Public

A light weight wrapper for pythons multiprocessing module that makes multiprocessing easy.

License

Notifications You must be signed in to change notification settings

root-11/mplite

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

93 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

mplite

Build status codecov Downloads Downloads PyPI version

A light weight wrapper for pythons multiprocessing module that makes multiprocessing easy.

In case anyone is looking for a very easy way to use multiprocessing with args and kwargs, here is a neat wrapper as mplite:

The test is also the showcase:

1. get the imports

from mplite import TaskManager, Task
import time

2. Create the function that each cpu should work on individually.

def f(*args, **kwargs):
    time.sleep(args[0])
    return args[0]/kwargs['hello']

2.1. I also add a function that will fail to illustrate that the TaskManager doesn't crash...

def broken(*args, **kwargs):
    raise NotImplementedError("this task must fail!")

3. create the main function you'd like to run everything from:

def main():
    args = list(range(10)) * 5
    start = time.time()
    
    with TaskManager() as tm:
        # add the first tasks
        tasks = [Task(f, *(arg/10,), **{'hello': arg}) for arg in args]

        print("an example of a tasks is available as string:\n\t", str(tasks[0]), '\n\t', repr(tasks[0]))

        results = tm.execute(tasks)   # this will contain results and tracebacks!
        
        end = time.time()
        print(f"did nothing for {end-start} seconds, producing {len(results)} results")
        print(f"hereof {len([result for result in results if isinstance(result, str) ])} had errors.")
        print(f"the rest where results: {[i for i in results if not isinstance(i,str)]}")
        
        # add more tasks to the SAME pool of workers:
        tasks = [Task(broken, *(i,)) for i in range(3)]
        results = tm.execute(tasks)
        print("More expected errors:")
        for result in results:
            print("expected -->", result)  

if __name__ == "__main__":
    main()

Expected outputs

an example of a tasks is available as string:
	 Task(f=f, *(0.0,), **{'hello': 0}) 
	 Task(f=f, *(0.0,), **{'hello': 0})

  0%|          | 0/50 [00:00<?, ?tasks/s]
  2%|▏         | 1/50 [00:00<00:07,  6.96tasks/s]
  4%|▍         | 2/50 [00:00<00:06,  7.75tasks/s]
  6%|▌         | 3/50 [00:00<00:05,  8.15tasks/s]
 14%|█▍        | 7/50 [00:00<00:03, 14.16tasks/s]
 18%|█▊        | 9/50 [00:00<00:02, 14.36tasks/s]
 24%|██▍       | 12/50 [00:00<00:02, 14.13tasks/s]
 32%|███▏      | 16/50 [00:01<00:01, 17.34tasks/s]
 38%|███▊      | 19/50 [00:01<00:01, 18.03tasks/s]
 42%|████▏     | 21/50 [00:01<00:01, 16.66tasks/s]
 46%|████▌     | 23/50 [00:01<00:01, 15.06tasks/s]
 52%|█████▏    | 26/50 [00:01<00:01, 17.60tasks/s]
 56%|█████▌    | 28/50 [00:01<00:01, 16.86tasks/s]
 62%|██████▏   | 31/50 [00:02<00:01, 16.72tasks/s]
 66%|██████▌   | 33/50 [00:02<00:00, 17.37tasks/s]
 70%|███████   | 35/50 [00:02<00:00, 17.72tasks/s]
 74%|███████▍  | 37/50 [00:02<00:00, 17.52tasks/s]
 80%|████████  | 40/50 [00:02<00:00, 19.88tasks/s]
 86%|████████▌ | 43/50 [00:02<00:00, 15.19tasks/s]
 90%|█████████ | 45/50 [00:02<00:00, 13.69tasks/s]
 94%|█████████▍| 47/50 [00:03<00:00, 14.46tasks/s]
 98%|█████████▊| 49/50 [00:03<00:00, 10.98tasks/s]
100%|██████████| 50/50 [00:03<00:00, 14.40tasks/s]

did nothing for 3.601374387741089 seconds, producing 50 results
hereof 5 had errors.
the rest where results: [0.1, 0.1, 0.0999..., 0.1, 0.1, 0.1, 0.1, 0.0999..., 0.0999..., 0.0999..., 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.0999..., 0.0999..., 0.0999..., 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.0999..., 0.0999..., 0.0999..., 0.1, 0.1, 0.1, 0.1, 0.1, 0.1, 0.0999..., 0.0999..., 0.0999..., 0.1, 0.1, 0.1, 0.1, 0.0999..., 0.0999..., 0.1, 0.1]

  0%|          | 0/3 [00:00<?, ?tasks/s]
100%|██████████| 3/3 [00:00<00:00, 80.66tasks/s]

More expected errors:

expected --> Traceback (most recent call last):
  File "d:\github\mplite\mplite\__init__.py", line 97, in execute
    return self.f(*self.args,**self.kwargs)
  File "d:\github\mplite\tests\test_basics.py", line 36, in broken
    raise NotImplementedError("this task must fail!")
NotImplementedError: this task must fail!

expected --> Traceback (most recent call last):
  File "d:\github\mplite\mplite\__init__.py", line 97, in execute
    return self.f(*self.args,**self.kwargs)
  File "d:\github\mplite\tests\test_basics.py", line 36, in broken
    raise NotImplementedError("this task must fail!")
NotImplementedError: this task must fail!

expected --> Traceback (most recent call last):
  File "d:\github\mplite\mplite\__init__.py", line 97, in execute
    return self.f(*self.args,**self.kwargs)
  File "d:\github\mplite\tests\test_basics.py", line 36, in broken
    raise NotImplementedError("this task must fail!")
NotImplementedError: this task must fail!

Note that tasks can't crash! In case of exceptions during task execution, the traceback is captured and the compute core continues to execute the next task.

How to test worker functions

Also, if you want to check that the inputs to the task are formed correctly, you can do the check from the interpreter, by calling .execute() on the task:

>>> t = Task(f, *(1,2,3), **{"this":42})
>>> t.execute()

How to handle incremental tasks

From version 1.1.0 it is possible to add tasks incrementally.

Let's say I'd like to solve the pyramid task where I add up all numbers

1+2  3+4  5+6  7+8  9+10
 =    =    =    =    = 
 3 +  7    11 + 15   19
   =         =       =
   10        26  +  19
   =             =
   10      +     45
           = 
          55

This requires that I:

  1. create a queue with 1,2,3,...,10
  2. add tasks for the numbers to be added pairwise
  3. receive the result
  4. when I have a pair of numbers submit them AGAIN.

Here is an example of what the code can look like:


def test_incremental_workload():
    with TaskManager() as tm:       
        # 1. create initial workload
        checksum = 55
        for a in range(1,10,2):
            t = Task(adder, a, a+1)
            print(t)
            tm.submit(t)
    
        # 2. create incremental workload
        a,b = None,None
        while True:
            result = tm.take()
            if result is None:
                if tm.open_tasks == 0:
                    break
                else:
                    continue
            
            if a is None:
                a = result
            else:
                b = result
            
            if a and b:
                t = Task(adder, a,b)
                print(t)
                tm.submit(t)
                a,b = None,None

        print(a,b,flush=True)
        assert a == checksum or b == checksum,(a,b,checksum)


Output:

Task(f=adder, *(1, 2), **{})
Task(f=adder, *(3, 4), **{})
Task(f=adder, *(5, 6), **{})
Task(f=adder, *(7, 8), **{})
Task(f=adder, *(9, 10), **{})
Task(f=adder, *(3, 7), **{})
Task(f=adder, *(11, 15), **{})
Task(f=adder, *(19, 10), **{})
Task(f=adder, *(26, 29), **{})
55 None

Use mplite wisely. Executing each tasks has a certain overhead associated with it. The fewer the number of tasks and the heavier (computationally) each of them the better.

Example with number of calls with a number of iterations in the call:

import multiprocessing
import time
from mplite import TaskManager, Task


def run_calcs_calls(mp_enabled=True, rng=50_000_000, calls=20, cpus=1):
    start = time.perf_counter()
    L = []
    if mp_enabled:
        with TaskManager(cpu_count=cpus) as tm:
            tasks = []
            for call in range(1, calls+1):
                tasks.append(Task(fun, *(call, rng)))
            L = tm.execute(tasks)
    else:
        for call in range(1, calls+1):
            res = fun(call, rng)
            L.append(res)

    task_times = [tm for res, tm in L]
    cpu_count = cpus if mp_enabled else 1
    cpu_task_time = sum(task_times)/cpu_count

    if mp_enabled:
        print('mplite - enabled')
    else:
        print('mplite - disabled')

    print('cpu_count: ', cpu_count)
    print(f'avg. time taken per cpu: ', cpu_task_time)
    end = time.perf_counter()
    total_time = end - start
    print('total time taken: ', total_time)
    print()
    return total_time, cpu_task_time, cpu_count


def fun(call_id, rng):
    # burn some time iterating thru
    start = time.perf_counter()
    t = 0
    for i in range(rng):
        t = i/call_id
    end = time.perf_counter()
    return t, end - start


def test_mplite_performance():    
    # change calls and range to see the knock on effect on performance
    print('========CALLS TEST===========')
    for cpus in [1, multiprocessing.cpu_count()]:
        for ix, (calls, rng) in enumerate([(10, 50_000_000), (2000, 50)], start=1):
            print('calls: ', calls, ', range: ', rng)
            total_time_mp_e, cpu_task_time_mp_e, cpu_count_mp_e = run_calcs_calls(True, rng, calls, cpus)
            total_time_mp_d, cpu_task_time_mp_d, cpu_count_mp_d = run_calcs_calls(False, rng, calls, cpus)
            artifacts = [cpus, calls, rng, total_time_mp_e, cpu_task_time_mp_e, cpu_count_mp_e, total_time_mp_d, cpu_task_time_mp_d, cpu_count_mp_d]
            if cpu_count_mp_e > cpu_count_mp_d:
                if ix == 1: # assert mplite is faster for less calls and heavier process
                    assert total_time_mp_e < total_time_mp_d, artifacts
            else:
                assert True

Output:

========CALLS TEST===========
calls:  10 , range:  50000000
mplite - enabled
cpu_count:  1
avg. time taken per cpu:  18.5264333
total time taken:  18.8809622

mplite - disabled
cpu_count:  1
avg. time taken per cpu:  18.912037
total time taken:  18.9126078

calls:  2000 , range:  50
mplite - enabled
cpu_count:  1
avg. time taken per cpu:  0.005216900000000357
total time taken:  0.490177800000005

mplite - disabled
cpu_count:  1
avg. time taken per cpu:  0.003248700000142435
total time taken:  0.003983699999999146

calls:  10 , range:  50000000
mplite - enabled
cpu_count:  12
avg. time taken per cpu:  3.410191883333333
total time taken:  4.978601699999999

mplite - disabled
cpu_count:  1
avg. time taken per cpu:  19.312383399999995
total time taken:  19.312710600000003

calls:  2000 , range:  50
mplite - enabled
cpu_count:  12
avg. time taken per cpu:  0.0005722500000000056
total time taken:  0.9079466999999966

mplite - disabled
cpu_count:  1
avg. time taken per cpu:  0.0038669999999427773
total time taken:  0.004872100000000046

Example with sleep time in each adder function:

import multiprocessing
import time
from mplite import TaskManager, Task


def run_calcs_sleep(mp_enabled, sleep=2, cpus=1):
    args = list(range(20))
    start = time.perf_counter()
    prev_mem = 0
    L = []

    if mp_enabled:
        with TaskManager(cpus) as tm:
            tasks = []
            for arg in args:
                tasks.append(Task(adder, *(prev_mem, arg, sleep)))
                prev_mem = arg
            L = tm.execute(tasks)
    else:
        for arg in args:
            res = adder(prev_mem, arg, sleep)
            L.append(res)
            prev_mem = arg

    end = time.perf_counter()

    cpu_count = cpus if mp_enabled else 1

    if mp_enabled:
        print('mplite - enabled')
    else:
        print('mplite - disabled')

    total_time = end - start
    print('cpu_count: ', cpu_count)
    print('total time taken: ', total_time)
    print()
    return total_time, cpu_count


def adder(a, b, sleep):
    time.sleep(sleep)
    return a+b


def test_mplite_performance():
    # change sleep times to see the knock on effect on performance
    print('========SLEEP TEST===========')
    for cpus in [1, multiprocessing.cpu_count()]:
        for ix, sleep in enumerate([2, 0.02, 0.01], start=1):
            print('sleep timer value: ', sleep)
            total_time_mp_e, cpu_count_mp_e = run_calcs_sleep(True, sleep, cpus)
            total_time_mp_d, cpu_count_mp_d = run_calcs_sleep(False, sleep, cpus)
            artifacts = [cpus, total_time_mp_e, cpu_count_mp_e, total_time_mp_d, cpu_count_mp_d]
            if cpu_count_mp_e > cpu_count_mp_d:
                if ix == 1:  # assert mplite is faster for longer sleep
                    assert total_time_mp_e < total_time_mp_d, artifacts
            else:
                assert True

Output:

========SLEEP TEST===========
sleep timer value:  2
mplite - enabled
cpu_count:  1
total time taken:  40.4222287

mplite - disabled
cpu_count:  1
total time taken:  40.006973200000004

sleep timer value:  0.02
mplite - enabled
cpu_count:  1
total time taken:  0.7628226999999868

mplite - disabled
cpu_count:  1
total time taken:  0.4116598999999894

sleep timer value:  0.01
mplite - enabled
cpu_count:  1
total time taken:  0.5629501999999889

mplite - disabled
cpu_count:  1
total time taken:  0.21054430000000934

sleep timer value:  2
mplite - enabled
cpu_count:  12
total time taken:  4.821827799999994

mplite - disabled
cpu_count:  1
total time taken:  40.011519899999996

sleep timer value:  0.02
mplite - enabled
cpu_count:  12
total time taken:  0.713870500000013

mplite - disabled
cpu_count:  1
total time taken:  0.41133019999998055

sleep timer value:  0.01
mplite - enabled
cpu_count:  12
total time taken:  0.6938743000000045

Ran 1 test in 192.739s
mplite - disabled
cpu_count:  1
total time taken:  0.20631170000001475



About

A light weight wrapper for pythons multiprocessing module that makes multiprocessing easy.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •  

Languages