Unlocking your CPU cores in Python (multiprocessing)

[@ArtML]

\o/ Yay! Long waited multiprocessing video! Always appreciate the humor in intros! :D Thanks a lot, I am on a path of making parallelization / multiprocessing to become a second nature in my coding - these videos help greatly! More topic suggestions: - Simple speed-ups using GPUs - Panda speedup by Dask - unlocking multiple cores - Numba, JAX and the overview of JIT compilers - Cython, and the most convenient (easy-to-use) wrappers for C++ implementations - All about Pickling, best practices/fastest ways to write picklers for novel objects

[@ajflink]

And GPU speedups without Nvidia.

[@tdug1991]

It's also worth noting that smaller chunk sizes may be better for unpredictably distributed job times, as one runner may randomly grab many expensive jobs, and lock the pool when the rest of the processes finish. Great video, as always!

[@unusedTV]

Your video is about two years late for me! I was working on a heat transfer simulation in Python where we had to compare hundreds of different input configurations. I knew about the GIL and multiprocessing in general outside of Python, but had to figure out myself how to get it to work. Eventually I settled on a multiprocessing pool and it worked wonders, because now we could run 32 simulations in parallel (Threadripper 1950x). Quick caveat that I don't hear you mention: a lot of processors have hyperthreading/SMT (intel/amd respectively), showing double the amount of cores in the task manager. In our case we found that spawning a process for each physical core provided better results than using all logical cores.

[@lawrencedoliveiro9104]

5:07 Threading is also useful for turning blocking operations into nonblocking ones. For example, asyncio provides nonblocking calls for reading and writing sockets, but not for the initial socket connection. Simple solution: push that part onto a separate thread.

[@michaellin4553]

The funny thing is, adding random noise is actually a useful thing to do. It's called dithering, and is used nearly everywhere in signal processing.

[@tommucke]

You would however apply it to the analog signal at about half the sampling rate in order of getting better results for the digital signal (and smoothen it with a capacitor afterwards). It makes no real sense to add it on the digital side which is the only thing python can do

[@gamma26]

@@tommucke Unless you're doing image processing and want to achieve that effect I suppose. Pretty niche tho

[@maxim_ml]

It can be used as data augmentation in training a speech recognition model

[@louisnemzer6801]

'I'm going to need those sound files with random noise added in my email inbox by five pm' 😅

[@SpeedingFlare]

That pool thing is so cool. I like that it spawns as many processes as there are cores available. I wish my work had more CPU bound problems

[@lawrencedoliveiro9104]

9:27 Actually, there is a faster way of sharing data between processes than sending picklable objects over pipes, and that is to use shared memory. Support for this is built into the multiprocessing module. However, you cannot put regular Python objects into shared memory: you have to use objects defined by the ctypes module. These correspond to types defined in C (as the name suggests): primitive types like int or float, also array and struct types are allowed. But avoid absolute pointers.

[@ЕвгенийКрасилов-о9о]

Aren't Managers a way to store shared python classes (via register)?

[@OutlawJackC]

Your explanation of the GIL makes so much more sense than other people :)

[@zemonsim]

This video was so helpful ! I recently converted my mass encryption script to use multiprocessing. To encrypt my dataset of 450 Mb of images, it went from an estimated 11 hours to just 10 minutes, doing the work at around 750 Kb per second.

[@michaelwang3306]

The clearest explanation on this topic that I have ever seen! Really nice!! Thanks for sharing!

[@etopowertwon]

Multiprocessing helped me a lot recently. I had a script that periodically loads lots tons of small XML from netshare, process them and save locally, single thread ran in 30 seconds, multiprocessed ran in about 6 seconds.

[@jesuscobos2201]

Love your videos. I usually watch all of them just for fun but this has enabled me to speed up a very heavy optimization for my science stuff. Ty for your dedication. I can ensure that it has real world implications :)

[@zlorf_youtube]

Me learns a new python thing... Starts using it in every fucking uncessary place. Feels good. Really good thing to talk about typical pitfalls.

[@superzcreationz2032]

This is what I was searching for.....very very very useful video...😊...that's why I subscribed you 😊

[@unotoli]

So well explained. One nice2have thing - quick tip on how to debug (see summary of time2process) most cpu-intensive tasks (functions, like wav transformation in this case).

[@knut-olaihelgesen3608]

You are actually the best at advanced python videos! Love them so much

[@jakemuff9407]

Great video! Maybe some more "real world" examples would be useful. Knowing that my code *could* be parallelized and actually parallelizing the code are two very different things. I've found that knowledge of multithreading in python does not translate to automatic code speed up. And of course no two problems are the same.

[@tytywuu]

I think it is more about doing experiments on asyncio/threading/multiprocessing on your own - everyone has different Python use cases

[@ibrahimaba8966]

multithreading is for io bound tasks, i use multiprocessing with zeromq to do some extensive image processing tasks!

[@chndrl5649]

Take crawling as example, it would be a huge time saver if you want crawl multiple words at a time

[@chndrl5649]

It all depends on how you can split your work.

[@v0xl]

python is not the right tool for high performance applicatons anyway

[@dlf_uk]

What are the benefits/drawbacks of this approach vs using concurrent.futures?

[@bersi3306]

Answer reside in the difference between concurrency and parallelism. When to use them also makes a lot of difference (here "CPU bounds" problems to solve with parallelism vs "I/O bounds" problems to solve with concurrency). You should also check (in the concurrent side) the difference between a Threaded function vs a coroutine.

[@JohnZakaria]

If numpy / scipy do the computations in C land, why don't they release the GIL and aquire it back when the computation is done? When writing a C++ module using pybind11, you have the option to release the Gil, granted that you are doing pure C++.

[@julius333333]

pretty sure it does

[@JohnZakaria]

@@julius333333 if it did, then threads would speed up the computation. Just like i/o calls that do release the GIL

[@jheins3]

Not an expert but far and based on your comment, you probably know 100x more than I do. With that being said I am going to speculate that the traditional behavior of numpy/scify follows a standard api call to an external C/C++ optimized library (a dll in windows). The API is essentially a function that initiates the c-land magic. For error handling and for how the GIL works, the function call waits to receive the output from c-land before handing it back. Because the API is essentially a function call, the GIL cannot be released till the function returns. Again that's a guess.

[@flybuy_7983]

THANK YOU MY BROTHER FROM ANOTHER COUNTRY AND ANOTHER FAMILY!!!

[@pamdemonia]

It's really interesting to see the threading results, (avg ~.2.5sec per file, but only 7.6 sec total). Cool.

[@Lolwutdesu9000]

While I'm not using multi-threading in my current work, I'll definitely save this video so I can one day return to it!

[@triola_3]

5:43 that sounded like the minecraft oof

[@daniilorekhov9191]

Would love to see a video on managing shared memory in multiprocessing scenarios

[@lawrencedoliveiro9104]

2:07 Remember that “I/O” can also include “waiting for a user to perform an action in a GUI”.

[@eldarmammadov7872]

liked the way to speak about all three modules asynchio, threading, multiprocessingin one vidoe

[@nocturnomedieval]

This is so good and clear. A must share. BTW, how these techniques relate to the case when you are using numba with option parallel=True?

[@riccardocapellino9078]

I tested this on my old code used for my thesis, which basically performs the same calculation hundreds of times with no I/O (calculates flows in a aircraft engine turbine stage). Took me 10 minutes to adjust the code and made it 40% FASTER

[@walterppk1989]

Brilliant video. Absolutely flipping gold

[@HitAndMissLab]

Thank you for summarising it so wall and in a such intelligible way.

[@user-zu1ix3yq2w]

i went down a rabbit hole, MP, numba, cython, pypy... The speedup people can get is insane.

[@nocturnomedieval]

Could you please help me to find the answer: numba with option parallel=True how it relatesto cores/threads/process? @D:

[@quillaja]

god that's so much easier than what i've been doing writing all the coordination junk around queue

[@piotradamczyk6740]

I was looking for this kind of lessons for years. please do more.

[@HomoSapiensMember]

really appreciate this, struggled understanding differences between map and imap...

[@mingyi456]

Please make a video about pickable objects and pickling, I would like to know more about it.

[@iUnro]

Hello. Can you explain what is the difference between multiprocessing and concurrent futures package? For me they look the same so I wonder why did you chose one over another.

[@m0Ray79]

And don't forget that pure Python is not an only option. Pyrex, which is translated to C/C++, opens even more broad bridges towards performance.

[@ДмитроПрищепа-д3я]

Why use Pyrex when there's Cython tho?

[@m0Ray79]

@@ДмитроПрищепа-д3я Pyrex is a python language superset. Cython is its translator. I metioned it in my videos.

[@ДмитроПрищепа-д3я]

@@m0Ray79 Cython is also a python superset tho. And no, Cython isn't a translator for Pyrex, it's a separate thing that was influenced by Pyrex back then. And Pyrex is kinda dead with its last stable release being 12 years old.

[@m0Ray79]

​@@ДмитроПрищепа-д3я The syntax and the whole idea was introduced in Pyrex, I'm still calling it the old name. Ok, let's say Pyrex became Cython. And the file extension is still .pyx.

[@lawrencedoliveiro9104]

2:49 Re “I say "CPU" a bunch but i actually mean "core"” --- remember that the term “core” for “CPU” was coined by Intel (and possibly other chipmakers) when they started putting the circuitry for multiple CPUs onto a single chip. The distinction isn’t really important, except that some proprietary server software from that time had licence fees that were calculated per-CPU, but somehow this was relaxed into “per-CPU-chip slot”. This way, if you had multiple CPUs in one chip, you didn’t have to pay as much as if the chips were in separate slots (which was quite common in servers in those days). Why did it matter? I guess to prevent a revolt by customers angry over licence fees ...

[@robertbrummayer4908]

Great video! Man, your videos are awesome. And every time I learn a little bit and get a little bit better, like you say :) Best wishes from Austria!

[@codewithjc4617]

This is great content, I’m a big fan of C++ and Python and this is just amazing

[@cjsfriend2]

You should do a video on using logging alongside with the multiprocessing pool

[@POINTS2]

Yes! Pool is the way to go. Definitely an improvement the threading and allows you to not have worry about the GIL.

[@jonathandawson3091]

Not always an improvement. A process costs a lot more overhead as he explained in the video. Other languages don't have the stupid GIL, hope that it's also removed from python someday.

[@sebastiangudino9377]

@@jonathandawson3091 It's a safety measure, it'll probably never be removed from python. If you really need "unsafe" threads you could probably just write your threaded function from c and inter-opt it with python. What her that's actually worth it is up you you but a lot of times it is not

[@lawrencedoliveiro9104]

The GIL is an integral part of reference-counting memory management. Getting rid of it completely means moving to Java-style pure garbage collection, where even the simplest of long-running scripts could end up consuming all the memory on your system. There is a project called “nogil”, which sets out to loosen some GIL restrictions a bit. That should give some useful speedups, without abandoning the GIL altogether.

[@Roule_n_Scratche]

Hey mCoding, could you make an video about Cython?

[@matejlinek287]

Wow, finally a mCoding video where I didn't learn anything new :-D Thank you so much James, now I can rest in peace :)

[@imbesrs]

You are the only person i keep video notis on for

[@imnotkentiy]

-It is the end -ha. All this time i've only been using 1/16 of my true power, behold -nani?!

[@marvalmoments2099]

Great teaching, simple and effective, I've using this Multiprocessing with my coroutin, my program is flying, lol

[@saketkr]

Awesome! This was so so helpful. Could you also make one about all these, i.e, asyncio, multi-threading, multi-processing and then workers, please?

[@goowatch]

You should preferably use per-core display to better show what you want to explain. Thanks for sharing your experience.

[@tobiasbergkvist4520]

On Linux/macOS you can use the fork-syscall to "send" things that can't be pickled, but only when using `Process`, and not when using `Pool`, since the process needs to get all the unpickleable data at startup, and can't receive it after it has started. The child processes inherits the parents memory with copy-on-write when using `fork`, meaning it only creates a copy of the memory if an attempt to modify it is made.

[@TheGmodUser]

Soo, what do you do of the object isn't pickable?

[@joshuaowen1941]

I love your videos man! Absolutely love them!

[@HerChip]

6:28 “blocks”; what do mean bij blocking? Without this result for loop, the processes dont seem to start: why is that? What if i dont need a for loop to do something with “results”?

[@ДмитроПрищепа-д3я]

Then you use a map instead of imap.

[@joshinils]

A video on how to figure out which pieces take the most time and optimizing for time would be great. what profilers are there for python, how do i use them, how do i use them right?

[@peterfisher3161]

"what profilers are there for python" Spyder and PyCharm have built in profilers.

[@joshinils]

@@peterfisher3161 ah, so I'd have to use those IDEs, not VS code... ok I'd rather have some cli solution or one that works with vs code.

[@replicaacliper]

Scalene is an amazing profiler especially on Linux

[@peterfisher3161]

@@joshinils Quickly looking up I found cProfile, which is a built-in and can be used from the terminal. Not much popped up on VS code.

[@jbusa5dimvzgkiik]

I've found yappi + gprof2dot to be really useful to find where asyncio applications are spending the CPU time.

[@caiomazzaferroadami]

Can somebody help me out? I'm trying to put some of the things he mentioned in the video in practice and ran through something weird. In 5:50, he uses the iterable object from pool.imap_unordered() to print the return arguments from 'etl' function (filename and duration) for each element in the sounds list. I'm trying to do something similar, but my function (equivalent to his 'etl') returns just one argument instead of two. However, when I try to print each element from that iterable object, my program just freezes and I have to kill it. I can't figure out what's wrong. Note: when I convert it into a list, i. e. list(pool.imap_unordered(fcn, iterable)), it seems to work fine for some reason.

[@mme725]

Nice, might play with this when I get off work later!

[@unperrier5998]

Can't wait for PEP 554 multiple interpreters to be mainline.

[@nathanthreeleaf4534]

What if instead of completing a process in each pool, the data that is returned from each "pool" needs to be stored somehow?

[@SkyFly19853]

Very useful for video game development.

[@austingarcia6060]

I was about to do something involving multithreading and this video appeared. Perfect!

[@felixfourcolor]

More videos on threading/asyncio please 😊

[@pedroarthurstudart1999]

5:42-5:43 and a reference to Mincraft damage taking.

[@ali-om4uv]

It would be great if you could show if this can be used for Ml hyperparamerer tuning and other Ml tasks.

[@adityaalmighty1]

The function inside Pool() does not read global variables. Can you please show a way to fix that? It has something to do with this Queue() class, isn't it? The Docs are a bit confusing

[@rabin-io]

Any chance for a follow-up using this inside of a Class? And compare it with pathos.multiprocessing?

[@pschweitzer524]

Now the question: would running threads within each multiprocess process be even faster?

[@MihaiNicaMath]

What is this, a CPU monitor window for ants? It needs to be at least 3 times as big! Joking aside, I enjoyed the video and learned something! The pitfalls are especially helpful. Thank you :)

[@EvanBurnetteMusic]

This is great! Thanks! Would love a guide on how to use shared memory with multiprocess. I've been optimizing a wordle solver that looks for five words with 25 unique letters as in the recent Stand Up Maths video. On my 8 core machine, each subprocess ends up using half a gig of memory! My data structure is a list of variable length sets. With pool I have to resort to pool.starmap(func, zip(argList1, argList2)) to pass all the data I need into each subprocess. Compared with my naive manual multiprocess implementation, the mp pool version is 30% slower. I'm hoping it can be faster with shared memory. Again, I really appreciate that you created an almost real world problem to demonstrate multiprocessing. It gave me the context I needed to implement this with my program.

[@volbla]

I tried using multiprocessing on my prime number sieve where each process have to write to the same array. It didn't really end up being faster (i'm probably bottlenecked by ram speed), but i did get the shared memory to work with numpy arrays. In your main process you do: shared_mem = SharedMemory(name = "John", create = True, size = #bytes) an_array = np.ndarray((#elements,), dtype = #type, buffer = shared_mem.buf) # Put your data in the array And in each subprocess you reference the memory by basically doing the same thing again. shared_mem = SharedMemory(name = "John") an_array = np.ndarray((#elements,), dtype = #type, buffer = shared_mem.buf) # Do something with the data In this case it was also useful to pass the process inputs through a Queue rather than function arguments. Then they only have to be instantiated once, even when consuming a lot of unpredictable data.

[@EvanBurnetteMusic]

@@volbla Thanks for the queue tip I will definitely be trying that out!

[@firefouuu]

I still not sure why the wavfile.read is able to run in parallel thread despite the GIL. Is it just because it's C code ? So, if for any reason this was written in pure python this would not work ?

[@aleale550]

Great video! You could do a follow up parallel computing video using Dask?

[@MithicSpirit]

3:24 huh, so the python standard doesn't require a GIL?

[@aadithyavarma]

Doesn't Python use pass by reference instead of pass by value, so does passing an large object to a method really matter here?

[@ДмитроПрищепа-д3я]

That's true, but here we pass that to another process, which happens by value (well, almost, it's pickled, passed as a binary data and then unpickled inside of another process).

[@aadithyavarma]

@@ДмитроПрищепа-д3я Since individual processes don't share memory, the data needs to be copied for each process. That's makes sense. Thanks!

[@czupryn0135]

If i have an lost of x,y coordinats and i need to calculate distance between each one of them. so to make it faster i cut the 1000 elements array into 5 samller 200 elemnts arrays. than how do i make fisrt core process 1 array, second core the 2 one and so on?

[@neelroshania7116]

This was awesome, thank you!

[@plays1361]

Great video, the program works great

[@volundr]

This is very useful, thank you

[@Kamel419]

I had to solve a complex problem similar to this and ended up needing to use a specific sequence of queues and workers to solve it. I think I ended up with 6 total workers, each with a "parent" worker flowing into it. I think it would be neat to showcase something like this

[@codedinfortran]

thank you. This made it all very clear.

[@ЭльмарИдрисов-г5э]

Are there any potential dangers/threats when using these methods? I understand that you can slow your program down instead of giving it speed, but besides that ? Any dangers to the computer itself or the data source (if it is coming from a database) ?

[@tehseensajjad1003]

Im learning stuff myself though here's what i can say about databases. Corrections/additions are welcome. Usually there are specialized drivers for doing stuff asynchronously with the database. Also ACID should take care of not ruining the database. As for damage to the computer, no. This is the intended way of doing things in a multi core processing unit. Dont be scared to push your computer. Altough the robot uprising hasnt happened yet, its safe to say, Computers are not humans.

[@ЭльмарИдрисов-г5э]

@@tehseensajjad1003 , thank you for your reply. I am planning to experiemnt with some of these methods for my projects. Let's see how many "time" gains it will give.

[@tehseensajjad1003]

@@ЭльмарИдрисов-г5э It can get very confusing trying to design your program around doing stuff parallel or concurrent at first, but it'll click one day. Good luck friend.

[@etopowertwon]

You don't want to do non-atomic operations that can leak outside. Like in SQL check first something with SELECT and INSERT it if it was not found "if not sql("SELECT Id FROM Table WHERE Table.Foo=1"): sql("insert into Table(Foo) values(1)")" Two processes can try to insert the same value to the table at the same time.

[@Darios2013]

Thank you for great explanation

[@ArgumentumAdHominem]

Great video. It would be super nice to have a worked example showing strong scaling. I have found that python multiprocessing is really not that great in terms of performance. Imagine the basic scenario: you need to compute a very expensive function of index 'i' between 0 and 1000. The function takes the same time for each index. There are no shared resources between processes. Naively one would expect the performance to scale as the inverse of the number of cores, but it is actually significantly worse, and I still don't fully understand why.

[@dinushkam2444]

Great video Very interesting stuff

[@GaryHost-qs9pg]

Very well done video. thank you

[@lakeguy65616]

I have a somewhat related question(s). I have a function where I open a file, perform a number of functions and then write the file to disk. without multiprocessing, it takes 1-2 minutes per file. I've modified my code to take advantage of the multi-cores on my pc. Its reduced the time by a factor of 3+. My problem is that its maxing out the CPU at 100% until the function finishes which means I can't use the pc for any other purpose while the multiprocessing is taking place. Heres my question. How can I reduce the work load on the CPU (even if it takes a little longer)? To process 100 files take at least 45 minutes. eventually I have 500+ files to process.... Any ideas? thank you!

[@therelatableladka]

from multiprocessing import Pool # Specify the number of cores to use num_cores = 4 # Change this to the desired number of cores with Pool(processes=num_cores) as pool: # Your code here Hope it helps

[@talhaibnemahmud]

Much needed video. I recently had to use multiprocessing for Image Processing & AI Game Assignment at the university. Although I used concurrent.futures.ProcessPoolExecutor() , this seems like a good option too. Maybe a comparison between these different options? 🤔

[@botondkalocsai5322]

Can async be used in conjuction with multiprocessing?

[@emilfilipov169]

OMG you used start and end time as part of the code?!?!?! In the meantime i get rejected on an interview because i didn't know how to write a decorator to do that same task.

[@ewerybody]

This is cool and all for relatively small python scripts. What if I have a UI (maybe Qt for Python) and want to kick off some work on a pool of processes. I wouldn't want these processes to load (or even execute) any of the UI code 🤔

[@mahmoudshihab]

I didn't quite understand pitfall number 3, when you showed: `items = [np.random.normal(size=10000) for _ in range(1000)] ` Why is this a pitfall? Also, for the fib demonstration... For some reason, fib took 1.35s vs nfib took 35.05s Even the normal implementation took less time than multiprocessing at 12.93s I even copied the fib and n_fib from your github to ensure that I wasn't doing something wrong But I can't seem to replicate your results

[@rohitathithya3964]

@7:59 bruh! and slapping the like, odd number of times , wow

[@anihilat]

Great Video!

[@hicoop]

Such a good video!

[@maheshcharyindrakanti8544]

took me a while due to mistake, but it works thanks

[@pamdemonia]

What is the difference between the concurrent.futures module and the multiprocessing module?

[@michael_p]

concurrent.futures is a high-level interface to both multiprocessing and multithreading.

[@replicaacliper]

I'm using Numba to optimize a program and I'm getting ~100% CPU usage. I want to run this program multiple times with independent parameters. In this case, would multiprocessing provide any real benefit over running the program one at a time?

[@SageBetko]

If Numba is already fully utilizing all CPU cores, then no, the overhead of adding Python’s multiprocessing into the mix will probably just slow things down.

[@s7gaming767]

This helped a lot thank you

[@anon_y_mousse]

Once upon a time, multi-processing required multiple full CPU's, so it's a very understandable speako. It might also show your age. Although, it might make for an interesting video to make a Beowulf cluster with RPi's and show how to program it to calculate something in parallel. Pi itself is obvious and easy, but perhaps how to do video encoding or 3D scene rendering would be a great fit.

[@harrytsang1501]

The best way to talk about multiprocessing and task scheduling is with RTOS. The important parts are in some 2000 lines of C and it's amazing for embedded systems

[@anon_y_mousse]

@@harrytsang1501 It might be pretty cool if he did a whole video series showing beginner methods in one and more advanced methods in an other. Using RPi OS with Python for the beginner series, RTOS and C for the more advanced.

[@necbranduc]

Awesome! What about using apply_async vs map?

[@AntonioZL]

Very useful. Thanks!

[@shikutoai]

Me: Audio Engineer You: 'Let's add random noise to a bunch of audio files' Me: AOWERIGLHAOLEIRGHAOPEIRGHSOLIERHGNBapoiewhftgliakrhegaopilerhgsireghsSGrfegplihjsoigjhsdfgoiHJPOIGHSOIDFHGSAPODRJfg Edit: Thanks for the intro to multiprocessing in Python