Sequential Exeuction, Multiprocessing, and Multithreading IO-Bound Tasks in Python

Python makes concurrency easy. It took less than an hour to add multiprocessing to my blog engine, First Crack, and I have used it often since. Everyone likes to call premature optimization the root of all evil, but architecting programs for concurrent execution from the start has saved me hundreds of hours in large data capture and processing projects. Color me a Knuth skeptic. This article compares sequential execution, multiprocessing, and multithreading for IO-Bound tasks in Python, with simple code samples along the way.

Terms #

First, terms. Most programs work from top to bottom. The next line runs after the last one finishes. We call these sequential. Adding multiprocessing to First Crack let the script use multiple cores to run multiple lines at the same time. Where the engine used to open a file, read its contents, close it, and then repeat those steps a thousand more times, it could now handle eight at once. Multithreading lives somewhere in the middle. These programs use a single core, but the processor forces small blocks — threads — to take turns. By pausing a block waiting to read a file so that another can make a network connection before coming back to the first, multithreading boosts efficiency and lowers runtime. The latter approaches are examples of concurrency, which — to make things easy — you can just think of as anything not sequential.

The rest of this article starts with a simple sequential script, after the section below, before moving on to much faster concurrent versions later. Each section also includes runtime analysis, so you can see just how big an impact concurrency can have.

Imports #

I excluded the import statements from the code in the following sections, for brevity’s sake. For those who want to follow along at home, make sure your script starts with these lines:

# Imports

from sys import argv # Capture command line parameters

from multiprocessing import Pool as CorePool # Multiprocessing

from multiprocessing.pool import ThreadPool # Multithreading

from time import sleep # Sleep

from math import ceil # Rounding

from datetime import datetime # Execution time

I will not use most of these functions for a while, but use all of them in time. If you do decide to follow along at home, you will need Python 3.7 or later.

Sequential Execution #

The first — and most common — approach to concurrency is to avoid it. Consider this simple script:


# Method: handle

# Purpose: Handle item.

# Parameters:

# - item: Item to handle (X)

# Return: 0 - Success, 1 - Fail (Int)

def handle(item):

    sleep(2)

    return 1 # Success



x = range(100)



t1 = datetime.now()

for each in x:

   handle(each)

t2 = datetime.now()

print("Sequential time: {}".format((t2-t1).total_seconds()))

The generic method handle does nothing — but by sleeping for two seconds, it simulates a long IO-bound task. I chose to simulate this type of task — as opposed to a CPU-bound one — because most of my recent projects have involved downloading and reading massive files. These types of jobs spend most of their time waiting on data from the network or an external hard drive, which requires little from the processor. In a real project, I might replace sleep(2) with code to download a web page, parse the HTML, and write to a file. For consistency across runs, I just use sleep here. x = range(100) creates a list from 0 to 99, and the for loop then calls handle one hundred times, once for each number.

We can model best-case runtime with the formula T = h * n + o, with T as total execution time, h as the amount of time handle takes to run, n as the number of times the method runs, and o as the overhead to initialize, operate, and exit the script. The script above should take just over 200 seconds to finish: T = 2 * 100 + o = 200 + o

The script ran for 200.17 seconds, which makes o — the overhead to initialize, operate, and exit the script, in seconds — equal to 0.17. Next, let’s add multiprocessing and see what changes.

Multiprocessing #

Consider the script below. handle remains unchanged. x = range(100) creates the same one hundred-item list from 0 to 99, but then Core_Orchestrator calls handle for each number 0 to 99.


# Method: handle

# Purpose: Handle item.

# Parameters:

# - item: Item to handle (X)

# Return: 0 - Success, 1 - Fail (Int)

def handle(item):

    sleep(2)

    return 1 # Success



# Method: Core_Orchestrator

# Purpose: Facilitate multiprocessing.

# Parameters:

# - input_list: List to farm out to cores (List)

# Return: True, All successful; False, At least one fail (Bool)

def Core_Orchestrator(input_list):

    pool = CorePool(processes=MAX_CORES)

    results = pool.map(handle, input_list)

    pool.close()

    pool.join()

    del pool

    return all(results)



x = range(100)



t1 = datetime.now()

Core_Orchestrator(x)

t2 = datetime.now()

print("Multiprocessing time: {}".format((t2-t1).total_seconds()))

Recall that multiprocessing lets the script use multiple cores to run multiple lines at the same time. My computer has eight cores, so Core_Orchestrator runs eight instances of handle at once. We can now model execution time with T = (h * n)/c + n/c*y + o.

Let me break this new formula down: we can represent the time to run handle n times with (h * n). (h * n)/c, then, becomes the time to run handle n times on c cores. Multiple cores introduce some overhead, though, which we can account for with n/c*y: the number of times the script will have to assign handle to a new core, n/c, times the unknown amount of time that takes, y. o is, again, the overhead to initialize, operate, and exit the script.

Assuming o remains constant, our new formula says we can expect the code above to take at least 25 seconds: T = (2 * 100)/8 + 100/8*y + 0.17 = 25.17 + 12.5y. Since we don’t have a value for y, though, we cannot say how far over. Let’s find out.

The script finishes in 0:00:32.09, which gives us a value of 0.55 for y. At this point, we have two extremes by which to judge performance: the sequential approach took 200.17 seconds, while the multiprocessor approach took 32.09 seconds. Let’s see if we can beat it with multithreading, next.

Multithreading #

In the script below, handle once again remains unchanged, x = range(100) creates the same one hundred-item list from 0 to 99, but then Thread_Orchestrator calls handle for each number 0 to 99. Thread_Orchestrator uses a max of eight threads.

Will this script match the performance of the sequential one, with T = h * n + o? It runs on a single core, after all. Or will it look more like the multiprocessed code, with execution time measured by T = (h * n)/c + n/c*y + o?


# Method: Thread_Orchestrator

# Purpose: Facilitate multithreading.

# Parameters:

# - input_list: List to farm out to threads (List)

# Return: True, All successful; False, At least one fail (Bool)

def Thread_Orchestrator(in_list):

    try:

        thread_pool = ThreadPool(8)

        results = thread_pool.map(handle, in_list)

        thread_pool.close()

        thread_pool.join()

        del thread_pool

        return all(results)

    except Exception as e:

        # print(e)

        return False



x = range(100)



t1 = datetime.now()

Thread_Orchestrator(x)

t2 = datetime.now()

print("Multithreading time: {}".format((t2-t1).total_seconds()))

This script finished in 32.08 seconds. Whether eight cores sleep for two seconds or a single core waits for eight threads to sleep for two seconds apiece, the same amount of time passes. As a result, the non-parallel multithreaded approach managed to match the parallel multiprocessing one. In general, the execution time for these two approaches will match for IO-bound tasks; it will not for CPU-bound tasks, though. If I had used a complex math operation that required many CPU cycles, the multiprocessing method would have split the work across eight cores, while a single one would have had to do all the work for each thread in the multithreaded code. The table below explains when to use these strategies.

BottleneckExampleOptimize with
IONetwork connection, file operationMultithreading
CPU Complex math problem, searchMultiprocessing

Given the simulated IO-bound task here, if the multiprocessing version ran just as fast as the multithreaded one, why bother with multithreading at all? Because the max number of cores a processor has is a hard physical limit, while the max number of threads is a logical one. I can never use more than eight cores, but I can use as many threads as my operating system will allow. In practice, I have found that limit hovers around 1,000 per process.

To answer my question from earlier, we can model multithreaded performance like we did multiprocessing1, with T = (h * n)/t + n/t*z + o — except with t as the number of threads used, and z as the overhead of assigning handle to a new thread. Using the execution time T of the last run, 32.08 seconds, we now have a value for z: 0.55. This formula also tells us that we can minimize T by increasing t toward its limit around 1,000. Let’s test this theory.

The script below uses 16 threads (t=16). According to our formula and assuming o and z remain constant, it should finish in about 16 seconds: T = 200/16 + 100/16(0.55) + 0.17 = 16.11


# Method: Thread_Orchestrator

# Purpose: Facilitate multithreading.

# Parameters:

# - input_list: List to farm out to threads (List)

# Return: True, All successful; False, At least one fail (Bool)

def Thread_Orchestrator(in_list):

    try:

        thread_pool = ThreadPool(16)

        results = thread_pool.map(handle, in_list)

        thread_pool.close()

        thread_pool.join()

        del thread_pool

        return all(results)

    except Exception as e:

        # print(e)

        return False



x = range(100)



t1 = datetime.now()

Thread_Orchestrator(x)

t2 = datetime.now()

print("Multithreading time: {}".format((t2-t1).total_seconds()))

The script finished in 16.07 seconds with 16 threads, and 2.15 seconds with 100. More threads over 100 could not make this faster, though, because the script only had 100 tasks to complete; more would just go unused. Is this the best we can do? No: if the multiprocessing code ran on a machine with 100 cores, each core would run handle once and all cores would run their instance at the same time; execution would take about 2 seconds, since handle takes 2 seconds. Are 2.15 seconds realistically the best we can do, though? Maybe; I don’t have a 100 core machine laying around — but perhaps we can get closer, by combining multiprocessing and multithreading.

Multiprocessing + Multithreading: The Blended Approach #

Getting multithreading and multiprocessing to work together took some work. I’ll walk you through the code first, then delve into the results.


# Global control variables

# MAX_CORES: Maximum number of cores

MAX_CORES = 8

multiprocessing.Pool() creates a handle through which the script delegates tasks to individual cores. This command uses multiprocessing.cpu_count() to define the number of available cores in the pool. In the virtual environment I wrote most of this article in, though, that function gave me incorrect results. Creating a variable MAX_PROCESSORS and then overriding multiprocessing.cpu_count() with it when instantiating the pool fixed the problem. Your mileage may vary.


# Method: Core_to_Thread_Orchestrator

# Purpose: Facilitate multiprocessing and multithreading.

# Parameters:

# - input_list: List to farm out to threads by core (List)

# Return: True, All successful; False, At least one fail (Bool)

def Core_to_Thread_Orchestrator(input_list):

    try:

        pool = CorePool(processes=MAX_CORES)

        n = ceil(len(x)/MAX_CORES)

        results = pool.map(Thread_Orchestrator, [list(input_list[i:i+n]) for i in range(0, len(input_list), n)])

        pool.close()

        pool.join()

        del pool, n

        return all(results)

    except Exception as e:

        # print(e)

        return False

Core_to_Thread_Orchestrator accepts an input list, conveniently named input_list, then creates a pool of cores. The line, results = pool.map(Thread_Orchestrator, [list(input_list[i:i+n]) for i in range(0, len(input_list), n)]), needs some extra explaining.

  1. Divide input_list into even sub-lists for each core. n = ceil(len(x)/MAX_CORES) uses ceil to make sure a list of 100 elements on an 8 core machine does not get split into 8 sub-lists with 12 elements each (int(100/8=12.5)=12). This would only account for 96 elements and orphan the last 4. For cases like this, ceil ensures 8 sub-lists are created with 13 elements each, where the last one has just 9. [list(input_list[i:i+n]) for i in range(0, len(input_list), n)] then splits input_list into even sub_lists such that each core will have about the same amount of work to do.
  2. Multithread the processing of each sub-list. pool.map hands each sub-list off to a different core’s multithreading function. This has two major benefits. First, this allows each core to supervise the multithreading of a fraction of input_list, rather than the entire thing. The system then has to create fewer threads per core, which means each core can spend less time pausing and resuming threads. In theory, this approach also multiplies the max number of possible threads: where one core might have tapped out at 1,000, 8 cores should manage 8,000 without issue. In practice, though, most systems limit thread count by process rather than by core; on my system, that limit hovers around 1,000.
  3. Capture success or failure for all cores. result becomes a list with a return value for each core.

return all(result) returns True if all processes succeeded, but False if any failed.


# Method: Thread_Orchestrator

# Purpose: Facilitate multithreading.

# Parameters:

# - input_list: List to farm out to threads (List)

# Return: True, All successful; False, At least one fail (Bool)

def Thread_Orchestrator(in_list):

    try:

        thread_pool = ThreadPool(len(in_list))

        results = thread_pool.map(handle, in_list)

        thread_pool.close()

        thread_pool.join()

        del thread_pool

        return all(results)

    except Exception as e:

        # print(e)

        return False

As we saw earlier, multithreading handles IO-bound tasks best with a thread for each task. Since Thread_Orchestrator now receives a variable number of tasks, it now calculates the appropriate number of threads to create with len(in_list). Again, using more threads than tasks would not improve runtime. results = thread_pool.map(handle, in_list) then assigns handle to a thread for each element in the input list, captures the results in an array just like the previous method, and returns True of all threads succeed. If any fail, Thread_Orchestrator returns False.


# Read number of items to generate test data set with from parameter.

# Default to 100.

if (len(argv) == 1):

    seed = 100

else:

    seed = int(argv[1])



# Expand the range to a list of values

# seed = 100 -> 100 element list

# seed = 500 -> 500 element list

x = range(seed)



# Print seed and results

print(f"Seed: {seed}")



# # Multiprocess

# t1 = datetime.now()

# if (Core_Orchestrator(x) == False):

#     print("-- Core orchestrator failed.")

# else:

#     t2 = datetime.now()

#     print("Multiprocessing time: {}".format((t2-t1).total_seconds()))



# Multithreading

t1 = datetime.now()

if (Thread_Orchestrator(x) == False):

    print("-- Thread orchestrator failed.")

else:

    t2 = datetime.now()

    print("-- Multithreading time: {}".format((t2-t1).total_seconds()))



# Multiprocessing + multithreading

t1 = datetime.now()

if (Core_to_Thread_Orchestrator(x)  == False):

    print("-- Processor to thread orchestrator failed.")

else:

    t2 = datetime.now()

    print("-- Multiprocessing and multithreading time: {}".format((t2-t1).total_seconds()))

The code above accepts a parameter for the number of times handle must run, and then records the runtime for the multithreaded method and the blended method. Even if one fails, the test continues. The snippet below tests the limits of both approaches by feeding the script values from 100 to 1,000 in increments of 100. It does this ten times, to lessen the impact of anomalous runs.

for k in {1..10}; do for i in {100..1000..100}; do python3 main.py $i >> $k".txt"; kill $(ps aux | grep python | awk '{print $2}') 2>/dev/null; done; done

I could have done this all in Python, but this approach does a few things for me. For one, it forces the script to initialize, execute, and exit for each set of tasks from 100 to 1,000. This lessens the chance of cache or memory usage impacting successive runs. kill $(ps aux | grep python | awk ‘{print $2}’) 2> /dev/null makes sure no Python processes stick around to interfere with those runs. Again, it also goes through this process ten times, to lessen the impact of anomalous runs. Together, these help give me as unbiased a picture of the script’s runtime as possible. Check out the results, tabled below:

Approach Tasks Run 1Run 2Run 3Run 4Run 5Run 6Run 7Run 8Run 9Run 10AVG
Blended100 3.567204 2.48169 2.469574 2.571985 2.495898 2.509916 2.487354 2.551873 2.520211 2.494282 2.6149987
Blended2002.62477 2.721776 2.623119 2.62792 2.666248 2.65166 2.656948 2.607229 2.581664 2.617885 2.6379219
Blended300 3.084136 2.766728 2.820711 2.718879 2.71363 2.730298 2.764282 2.694365 2.71308 2.710206 2.7716315
Blended400 3.834809 3.039956 3.025324 2.826441 2.831418 2.902106 2.811155 2.893164 2.824065 2.815037 2.9803475
Blended500 3.412031 3.085576 3.107757 2.994405 2.970532 3.066721 2.962216 3.010268 2.919665 2.920715 3.0449886
Blended600 3.920105 3.276128 3.20756 3.191376 3.24736 3.18067 3.095356 3.182228 3.177074 3.12045 3.2598307
Blended700 3.893453 3.338353 3.45412 3.307333 3.269286 3.375174 3.416062 3.189951 3.309285 3.223167 3.3776184
Blended800 4.396074 3.409177 3.489747 3.441516 3.416976 3.33599 3.333954 3.409126 3.298936 3.52455 3.5056046
Blended900 4.138501 3.736235 3.642134 3.608156 3.699388 3.624607 3.776716 3.607106 3.60067 3.532703 3.6966216
Multithreaded 100 2.530468 2.678918 2.375514 2.669632 2.716273 2.433133 2.633597 2.528427 2.689116 2.529757 2.5784835
Multithreaded 200 2.646297 2.301428 2.220176 2.269561 2.33174 2.266194 2.251354 2.290965 2.278389 2.307648 2.3163752
Multithreaded 300 2.516477 2.406957 2.401755 2.533726 2.390161 2.862111 2.38876 2.384127 2.389991 2.398563 2.4672628
Multithreaded 400 2.482953 2.546255 2.676493 2.495154 2.58571 2.522376 2.571323 2.506227 2.574394 2.468326 2.5429211
Multithreaded 500 3.363701 2.85764 2.775013 2.604138 2.645682 2.601216 2.622577 2.702628 2.785129 2.667436 2.762516
Multithreaded 600 2.987607 2.724781 2.882752 2.681507 2.788063 3.160047 2.780519 3.312241 3.519288 2.788334 2.9625139
Multithreaded 700 3.457589 3.197889 2.950593 2.88573 2.986151 2.9273 2.890827 3.018946 2.894577 3.189955 3.0399557
Multithreaded 800 3.291344 3.208601 3.031519 2.981974 2.979717 6.996563 3.098343 3.908249 3.083194 7.007135 3.9586639
Multithreaded 900 3.532445 3.201465 3.164539 3.478344 5.107173 3.211502 3.120932 3.690724 3.577692 3.147345 3.5232161

The graph below visualizes execution time as a function of tasks, from 100 to 900. After 900, the system refused to spawn new threads; the dotted lines predict execution time beyond that point. y = 0.0237x2 - 0.0655x + 2.4838 models the multithreading method with a R² value of 0.83, and y = 0.0038x2 + 0.1019x + 2.4676 models the blended method with a R² value of 0.99.

Execution time as a function of tasks

The multithreaded method’s runtime consistently spikes with 800 tasks. Interestingly, normalizing the average runtime for 800 tasks to fall between 700’s and 900’s changes the trendline function from y = 0.0237x2 - 0.0655x + 2.4838 to y = 0.0185x2 - 0.0481x + 2.4838, and causes the R² value to jump from 0.83 to 0.96. Check out that graph below.

Normalized execution time as a function of tasks

Conclusions and Takeaways #

To return to my question from earlier, are 2.15 seconds the best we can do? Recall that Thread_Orchestrator blew through 100 simulated IO-bound tasks in 2.15 seconds, using 100 threads. Over ten runs, it averaged 2.16 seconds; the blended multiprocessed + multithreaded method, on the other hand, averaged 2.44 seconds over 10 runs. To answer my question from earlier, then, 2.15 seconds are the best we can do. Multithreading wins for IO-bound tasks.

As the number of IO-bound tasks grows, that eventually changes. The multithreaded method’s execution time stays below the blended method’s from 100 to 900 tasks, but the former grows faster than the latter. On a system that permits a process to spawn more than 1,000 threads, the blended approach will begin to win out when processing over 1,000 IO-bound tasks. The table below summarizes when to use multithreading, multiprocessing, or a mix of both.

Bottleneck Example Tasks Optimize with
CPU Complex math problem, search Any Multiprocessing
IO Network connection, file operation < 1,000 Multithreading
IO Network connection, file operation > 1,000 Multiprocessing + multithreading

Use this table to choose an approach, and the scripts above to make quick work of even large jobs. Multi-core, multithreaded architectures mean no one should have to suffer through painful sequential execution anymore. Python makes concurrency easy, so take advantage of it.

 I understand that T = (h * n)/t + n/t*z + o implies simultaneous execution, which is correct when using multiprocessing but not when using multithreading. Multithreaded programs run on a single core. Although the processor pauses and resumes threads so fast that it gives the impression of parallel execution, they do not execute in parallel. In this scenario, though, this is effectively a meaningless distinction. Multithreaded IO-bound tasks are essentially indistinguishable from multiprocessed ones, given an equal number of threads and cores.

Permalink.