One thing I'm curious about here is the operational impact.<p>In production systems we often see Python services scaling horizontally
because of the GIL limitations. If true parallelism becomes common,
it might actually reduce the number of containers/services needed
for some workloads.<p>But that also changes failure patterns — concurrency bugs,
race conditions, and deadlocks might become more common in
systems that were previously "protected" by the GIL.<p>It will be interesting to see whether observability and
incident tooling evolves alongside this shift.
> observability tooling for Python evolving<p>As much as I dislike Java the language, this is somewhere where the difference between CPython and JVM languages (and probably BEAM too) is hugely stark. Want to know if garbage collection or memory allocation is a problem in your long running Python program? I hope you're ready to be disappointed and need to roll a lot of stuff yourself. On the JVM the tooling for all kinds of observability is immensely better. I'm not hopeful that the gap is really going to close.
> If true parallelism becomes common, it might actually reduce the number of containers/services needed for some workloads<p>Not by much. The cases where you can replace processes with threads and save memory are rather limited.
This is surely why Facebook was interested in funding this work. It is common to have N workers or containers of Python because you are generally restricted to one CPU core per Python process (you can get a bit higher if you use libs that unlock the GIL for significant work). So the only scaling option is horizontal because vertical scaling is very limited. The main downside of this was memory usage. You would have to load all of your code and libraries N types and in-process caches would become less effective. So by being able to vertically scale a Python process much further you can run less and save a lot of memory.<p>Generally speaking the optimal horizontal scaling is as little as you have to. You may want a bit of horizontal scaling for redundancy and geo distribution, but past that vertically scaling to fewer larger process tend to be more efficient, easier to load balance and a handful of other benefits.
> The main downside of this was memory usage. You would have to load all of your code and libraries N types and in-process caches would become less effective.<p>You can load modules and then fork child processes. Children will share memory with each other (if they need to modify any shared memory, they get copy-on-write pages allocated by the kernel) and you'll save quite a lot on memory.
Yes, this can help a lot, but it definitely isn't perfect. Especially since CPython uses reference counting it is likely that many pages get modified relatively quickly as they are accessed. Many other GC strategies are also pretty hostile to CoW memory (for example mark bits, moving, ...) Additionally this doesn't help for lazy loaded data and caches in code and libraries.
For big things the current way works fine. Having a separate container/deployment for celery, the web server, etc is nice so you can deploy and scale separately. Mostly it works fine, but there are of course some drawbacks. Like prometheus scraping of things then not able to run a web server in parallel etc is clunky to work around.<p>And for smaller projects it's such an annoyance. Having a simple project running, and having to muck around to get cron jobs, background/async tasks etc. to work in a nice way is one of the reasons I never reach for python in these instances. I hope removing the GIL makes it better, but also afraid it will expose a whole can of worms where lots of apps, tools and frameworks aren't written with this possibility in mind.
A lot of that has already been solved for by scaling workers to cores along with techniques like greenlets/eventlets that support concurrency without true multithreading to take better advantage of CPU capacity.
But python can fork itself and run multiple processes into one single container. Why would there be a need to run several containers to run several processes?<p>There's even the multiprocessing module in the stdlib to achieve this.
Threads are cheap, you can do N work simultaneously with N threads in one process, without serialization, IPC or process creation overhead.<p>With multiprocessing, processes are expensive and work hogs each process. You must serialize data twice for IPC, that's expensive and time consuming.<p>You shouldn't have to break out multiple processes, for example, to do some simple pure-Python math in parallel. It doesn't make sense to use multiple processes for something like that because the actual work you want to do will be overwhelmed by the IPC overhead.<p>There are also limitations, only some data can be sent to and from multiple processes. Not all of your objects can be serialized for IPC.
Forking and multi threading do not coexist. Even if one of your transitive dependencies decides to launch a thread that’s 99% idle, it becomes unsafe to fork.
Im curious as to the down votes on this. It's absolutely true, and when I was maintaining a job runner daemon that ran hundreds of thousands of who knows what Python tasks/jobs a day on some shared infra with arbitrary code for a certain megacorp from 2016-2020 or so, this was one of insidious and ugly failure modes to go debug and handle. The docs really make it sound like you can mix threading and multiprocessing but you can never really completely ensure that threading and then bare fork will ever be safe, period. It's really irritating that the docs would have you believe that this is OK or safe, but is in keeping with the Python philosophy of trying to hide the edge of the blade you're using until it's too late and you've cut the shit out of yourself.
Why is it unsafe?
Fork-then-thread works, does it not?
If you have enough discipline to make sure you only create threads after all the forking is done, then sure. But having such discipline is harder than just forbidding fork or forbidding threads in your program. It turns a careful analysis of timing and causality into just banning a few functions.
But not the reverse, if its a bare fork and not strictly using basically mutex and shared resource free code (which is hard), and there's little or no warning lights to indicate that this is a terrible idea that fails in really unpredictable and hard to debug ways.
Should have funded the entire GIL-removal effort by selling carbon credits. Here's an industry waiting to happen: issue carbon credits for optimizing CPU and GPU resource usage in established libraries.
That reminded me of how back in 2008 I removed the GIL from Python to run thousands Python modules in 10,000 threads. We were fighting for every clock cycle and byte and it worked. It took 20 years for the GIL to be removed and become available to the public.
> Similarly, workloads where threads frequently access and modify the same objects show reduced improvements or even degradation due to lock contention.<p>Perhaps I'm stating the obvious, but you deal with this with lock-free data structures, immutable data, siloing data per thread, fine-grain locks, etc.<p>Basically you avoid locks as much as possible.
Our experience on memory usage, in comparison, has been generally positive.<p>Previously we had to use ProcessPoolExecutor which meant maintaining multiple copies of the runtime and shared data in memory and paying high IPC costs, being able to switch to ThreadPoolExecutor was hugely beneficially in terms of speed and memory.<p>It almost feels like programming in a modern (circa 1996) environment like Java.
Swapping ProcessPoolExecutor for ThreadPoolExecutor gives real memory and IPC wins, but it trades process isolation for new failure modes because many C extensions and native libraries still assume the GIL and are not thread safe.<p>Measure aggressively and test under real concurrency: use tracemalloc to find memory hotspots, py-spy or perf to profile contention, and fuzz C extension paths with stress tests so bugs surface in the lab not in production. Watch per thread stack overhead and GC behavior, design shared state as immutable or sharded, keep critical sections tiny, and if process level isolation is still required stick with ProcessPoolExecutor or expose large datasets via read only mmap.
Might be worth noting that this seems to be just running some tests using the current implementation, and these are not necessarily general implications of removing the GIL.
Sections 5.4 and 5.5 are the interesting ones.<p>5.4: Energy consumption going down because of parallelism over multiple cores seems odd. What were those cores doing before? Better utilization causing some spinlocks to be used less or something?<p>5.5: Fine-grained lock contention significantly hurts energy consumption.
I'm not sure of the exact relationship, but power consumption increases greater than linear with clock speed. If you have 4 cores running at the same time, there's more likely to be thermal throttling → lower clock speeds → lower energy consumption.<p>Greater power draw though; remember that energy is the integral of power over time.
Running a program either on 1 core or on N cores, ideally does not change the energy.<p>On N cores, the power is N times greater and the time is N times smaller, so the energy is constant.<p>In reality, the scaling is never perfect, so the energy increases slightly when a program is run on more cores.<p>Nevertheless, as another poster has already written, if you have a deadline, then you can greatly decrease the power consumption by running on more cores.<p>To meet the deadline, you must either increase the clock frequency or increase the number of cores. The latter increases the consumed energy only very slightly, while the former increases the energy many times.<p>So for maximum energy efficiency, you have to first increase the number of cores up to the maximum, while using the lowest clock frequency. Only when this is not enough to reach the desired performance, you increase the clock frequency as little as possible.
By running more tasks in parallel across different cores they can each run at lower clock speed and potentially still finish before a single core at higher clock speeds can execute them sequentially.
5.4 is the essential reason why multithreading has become the main method to increase CPU performance after 2004. For reaching a given level of performance, increasing the number of cores at the same clock frequency needs much less energy than increasing the clock frequency at the same number of cores.<p>5.5 depends a lot on the implementation used for locks. High energy consumption due to contention normally indicates bad lock implementations.<p>In the best implementations, there is no actual contention. A waiting core only reads a private cache line, which consumes very little energy, until the thread that had hold the lock immediately before it modifies the cache line, which causes an exit from the waiting loop. In such implementations there is no global lock variable. There is only a queue associated with a resource and the threads insert themselves in the queue when they want to use the shared resource, providing to the previous thread the address where to signal that it has completed its use of the resource, so the single shared lock variable is replaced with per-thread variables that accomplish its function, without access contention.<p>While this has been known for several decades, one can still see archaic lock implementations where multiple cores attempt to read or write the same memory locations, which causes data transfers between the caches of various cores, at a very high power consumption.<p>Moreover, even if you use optimum lock implementations, mutual exclusion is not the best strategy for accessing a shared data resource. Even optimistic access, which is usually called "lock-free", is typically a bad choice.<p>In my opinion, the best method of cooperation between multiple threads is to use correctly implemented shared buffers or message queues.<p>By correctly implemented, I mean using neither mutual exclusion nor optimistic access (which may require retries), but using dynamic partitioning of the shared buffers/queues, which is done using an atomic fetch-and-add instruction and which ensures that when multiple threads access simultaneously the shared buffers or queues they access non-overlapping ranges. This is better than mutual exclusion because the threads are never stalled and this is better than "lock-free", i.e. optimistic access, because retries are never needed.
Can’t it just profile them and pick the right one accordingly?
Title shortened - Original title:<p>Unlocking Python’s Cores: Hardware Usage and Energy Implications of Removing the GIL<p>I am curious about the NumPy workload choice made, due to more limited impact on CPython performance.
From [2603.04782] "Unlocking Python's Cores: Hardware Usage and Energy Implications of Removing the GIL" (2026) <a href="https://arxiv.org/abs/2603.04782" rel="nofollow">https://arxiv.org/abs/2603.04782</a> :<p>> Abstract: [...] <i>The results highlight a trade-off. For parallelizable workloads operating on independent data, the free-threaded build reduces execution time by up to 4 times, with a proportional reduction in energy consumption, and effective multi-core utilization, at the cost of an increase in memory usage. In contrast, sequential workloads</i> do not benefit from removing the GIL <i>and instead show a 13-43% increase in energy consumption</i>
[dead]
[flagged]
Thanks ChatGPT, good of you to let us know.
There are so many ChatGPT responses in this thread, it’s giving me a headache.
I'm curious what makes that obviously llm? As far as I can tell it was a short and fairly benign statement with little scope to give away llm-ness?
It's just the equivalent of that one student restating what the teacher just said with no added value
Just as bad if it's human. No information has been shared. The writer has turned idle wondering into prose:<p>> Once threads actually run concurrently, libraries (which?) that never needed locking (contradiction?) could (will they or won't they?) start hitting race conditions in surprising (go on, surprise me) places.
It was an essentially pointless platitude about the GIL from a very new account not really related to the article, and all comments from this account are the same: top level comments with lots of em-dashes that are just a vague piece of pablum somewhat related to the subject. If it was just this comment, sure, it could be possible it's a rather uninteresting human. But given the history, this account is pure AI slop.
The obvious solution is to require libraries that are no-GIL safe to declare that, and for all other libraries implicitly wrap them with GIL locks.
Academics are cooked if Python ever gets good, it will be invaded by real developers who use JS.<p>Thankfully that piece of shit language is not only slow but has no consumer runtime environment like JS' browser, Swift's iOS, etc. and will slowly go the way of Ruby, PHP, etc.<p>Python devs need to abandon that trash and learn Node, or make it better and I'll do Python IDGAF
I have a suspicion that this paper is basically a summary with some benchmarks, done with LLMs.
Your suspicion could have easily been cleared by reading the paper.<p>If you're short on time: the paper reads a bit dry, but falls in the norm for academic writing. The github repo shows work over months on 2024 (leading up to the release of 3.13) and some rush on Dec 2025 to Jan 2026, probably to wrap things up on the release of this paper. All commits on the repo are from the author, but I didn't look through the code to inspect if there was some Copilot intervention.<p>[0] <a href="https://github.com/Joseda8/profiler" rel="nofollow">https://github.com/Joseda8/profiler</a>
> Across all workloads, energy consumption is proportional to execution time<p>Race-to-idle used to be the best path before multicore. Now it's trickier to determine how to clock the device. Especially in battery powered cases. This is why all modern CPU manufacturers are looking into heterogeneous compute (efficiency vs performance cores).<p>Put differently, I don't think we should be killing ourselves over this at software time. If you are actually concerned about the impact on raw energy consumption, you should move your workloads from AMD/Intel to ARM/Apple. Everything else would be noise compared to this.
Programs whose performance is dominated by array operations, as it is the case for most scientific/technical/engineering applications, achieve a much better energy efficiency on the AMD or Intel CPUs with good AVX-512 support, e.g. Zen 5 Ryzen or Epyc CPUs and Granite Rapids Xeons, than on almost all ARM-based CPUs, including on all Apple CPUs (the only ARM-based CPUs with good energy efficiency for such applications are made by Fujitsu, but they are unobtainium).<p>So if you want maximum energy efficiency, you should choose well your CPU, but a prejudice like believing that ARM-based CPUs are always better is guaranteed to lead to incorrect decisions.<p>The Apple CPUs have exceptional and unmatched energy efficiency in single-thread applications, but their energy efficiency in multi-threaded applications is not better than that of Intel/AMD CPUs made with the same TSMC CMOS fabrication process, so Apple can have only a temporary advantage, when they use first some process to which competitors do not have access.<p>Except for personal computers, the energy efficiency that matters is that of multi-threaded applications, so there Apple does not have anything to offer.
this is a very silly take. cpu isa is at most a 2x difference, and software has plenty of 100x differences. most of the difference between Windows and macos isn't the chips, OS and driver bloat is a much bigger factor
CPU ISA is at most a 2x difference for programs that use only the general-purpose registers and operations.<p>For applications that use vector or matrix operations and which may need some specific features, it is frequent to have a 4x to 10x better performance, or even more than this, when passing from a badly-designed ISA to a well-designed ISA, e.g. from Intel AVX to Intel AVX-512.<p>Moreover, there are ISAs that are guilty of various blunders, which lower the performance many times. For instance, if an ISA does not have rotation instructions, an application whose performance depends a lot on such operations may run up to 3x slower than on an ISA with rotation instructions<p>Even greater slow-downs happen on ISAs that lack good means for detecting various errors, e.g. when running on RISC-V a program that must be reliable, so it has to check for integer overflows.