If your algorithm does a ton of small allocations to the point where the allocator is the bottleneck, you're already doing it wrong. The allocator necessarily comes with a lot of overhead because it needs to accommodate diverse use cases, avoid fragmentation, and ideally, implement a variety of security checks. If you're doing something alloc-intensive, you're probably allocating and freeing a lot of identical structures and you'd be better off grabbing some continuous memory and managing that yourself in a task-specific way.
But the reality is that almost no one actually cares about performance because compute is cheaper than expertise and labor, at least in the short haul. Everything is getting more bloated and slower and we just compensate by adding CPU cores, gigabytes and gigahertz.
No, musl's allocator is just bad even in completely normal programs, and it is especially awful if you are using even two threads much less a lot of them. It has no TLABs or arenas. It has a single global mutex over alloc/free paths. It does syscalls underneath that lock (mmap) meaning the few fast paths it has are rarely taken under contention and have to fall back to futex wakes, so even 2 threads with minor contention and allocation rate will have visible wait points in profiles, stuck waiting for the allocator. It returns mapped memory to the OS very eagerly when a size class is empty, so even single allocs followed by a single free can cause thrashing as it mmaps/unmmaps things repeatedly for a size class over and over. Etc. You quite literally have to limit your thread count when using musl, because it will tank the performance of actually highly threaded programs that can scale with core count, even at very modest allocation rates and small working set sizes.
Its string routines and memory copy routines are also similarly bad, as the article alludes to. They are just naive loops with nearly no optimization. These are not small insignificant functions where using them is "doing it wrong", they are the backbone of vast amounts of code and can be made multiple times faster. You can similarly see string routines pop up in profiles all the time in musl builds in my experience. And unlike the memory allocator these cannot be "fixed" systematically across the application at link time, so you are stuck with it.
Real programs have to often do things like allocate memory and use multiple threads and process strings. People have been optimizing these things for decades, there is vast amounts of prior art, the musl developers simply did not do so because they prioritize simplicity over nearly everything else (from what I can tell) including performance.
Think about it this way because the issue isn’t specific to allocators: it’s pretty good in general but can often be beaten if you have special understanding of what you need to do. That’s OK.
You can buy cars and trucks that are optimized for driving on freeways and residential streets carrying stuff people often carry. But then there are special vehicles like fork lifts and such that are kinds of large special cases. And then there are weirdo specialised vehicles that have four wheels but are rare and their users can’t live without them.
Languages like C++ let you plug in special allocators if you want. But most people don’t. Some, like HFT people do crazy headstands to avoid slow allocations. I don’t ever want to do that but if they want to, why not. I don’t think they complain that the default case doesn’t fit their needs!
"But the reality is that almost no one actually cares about performance because compute is cheaper than expertise and labor, at least in the short haul."
Doesn't have to stay that way, with hardware prices soaring and development cost allegedly in free fall.
People have such different perspectives. 26% slower does not sound "terrible" to me; it sounds like quite a reasonable price one might choose to pay for the convenience musl offers. If musl's allocator were 2.6x slower, I might call that "not so great"... but in order to qualify as "terrible" I think the difference would have to be an order of magnitude!
The 26% number at the top of the article is from using mimalloc (which is a high performance allocator, at least as fast as the glibc allocator) + musl for some task, and the slowdown is coming from (probably) slow musl implementations of memcpy/memset. The musl allocator is even worse.
Yeah, doing compute-heavy work a couple of jobs ago, we tried small images with musl, and the default allocator was a catastrophe: 75%+ slowdowns for our real life tasks. Even with a better allocator, we were way better off with the larger image.
> the slowdown is coming from (probably) slow musl implementations of memcpy/memset.
It's wild that such a fundamental piece of code (you can't really implement operation on structs without those) is library-supplied. I wish compilers would just have something like __builtin_memcpy and __builtin_memset, and provided some highly optimized, specialist-crafted assembly in those, instead of having to inline the library code and hopefully be able to optimize it.
Clang and GCC do provide these, and automatically use them in many situations (particularly small copies). But c-libraries can actually do it better in many cases, especially for large copies.
Glibc, for example, has perhaps ten different implementations of memcpy just for x86. The compiler certainly could provide all that, but the next step is harder:
glibc automatically dispatches to the proper one at runtime based on the actual microarchitecture that the binary is running on. You pay the extra dispatch cost once, but all of non-inline function call cost every time. This is what allows distros to compile to a nice baseline architecture, but still get near-optimal memcpy performance on many more architectures than a single inline instance could possibly give. These differences matter.
And it does it for not just memcpy, but half-a-dozen other extremely performance sensitive library functions, like strcpy and so on.
Inlining works very much against this strategy. If you can guarantee that the target microarch never changes, then it isn't a good one. But that is somewhat unusual for everyone but those who build their own binaries to run on a single class of machines forever.
Worse, inlining the really high performance versions of these ends up being terrible from a code size perspective, because they are often hundreds of instructions, which can have bad caching effects. And once you amortize the function-call cost over many iterations of the loop, it isn't so expensive to call out to the library.
Anyway, just some additional considerations to think about.
Even if the attempt is inside of a function called memcpy() which contains no code other than your copy loop, and links with priority over the libc implementation! (as all embedded firmware engineers learn at some point in their journey)
Maybe you're being sarcastic, but I'm pretty sure clang + gcc do offer these.
The problems at first glance :
- Not having control over the implementation detail of the interface that your library provides is probably not wise. Sounds like a lot of bad bug reports and edge cases that you have no control over.
The 26% slower appears to be for their whole application, not just the allocator. For some parts of the application to make the whole this much slower it must mean that those parts are quite a lot slower, likely much more than 2x.
Moreover the 26% is with mimalloc, with musl's allocator it's 144%, so there are likely other parts that are slower (likely the memcpy implementation)
Ops here, I think if you NEED that convenience, sure, rock with MUSL BUT I also see a ton of devs crowing about using MUSL on my 128GB x86 Kubernetes hosts. I have plenty of Disk Space, you can ship glibc based container.
I feel with Rust I try and avoid re-allocations in most cases anyway, so I'm not sure that musl's allocator being slow would significantly affect performance (though I haven't benchmarked it). I feel like part of the appeal of Rust is that you can do imperatively-style mutation-heavy code comparatively risk-free, so despite me normally being the "Functional Programming Nerd", I generally write Rust in a style that's a bit closer to C.
I use musl for my Rust stuff because I have noticed that for the stuff I write it appears to have a lower memory footprint; since a lot of what I do is IO-bound anyway, I care more about using less memory than raw performance.
Musl's allocator being awful is pretty well known, though mostly in that it's absolutely awful in multithreaded context. TFA points out that musl has a bunch of other noticeably slower functions, which is less well known (though they're also slower by a smaller factor, and they don't worsen as your parallelism increases).
Yeah... I've recently had a chance to compare how fgets is implemented in both GNU libc and musl, and, well. With glibc, it was a challenge to even find where the fgets's code actually is.
I haven’t dug into why, but for unknown-linux builds on x86, Rust binaries have been substantially smaller on musl than standard dynamic linking to glibc, for me. No idea if I’m doing something wrong or if the handful of cases I tried were all special in some way.
You can do FROM scratch, and use still glibc; it’s just that you need to copy more than one file. I don’t really understand if you are already dealing with images why you still need the image to contain a single file.
If this specific use case is of high interest to you and you have some available bandwidth, contributing to it, maybe becoming a maintainer, and eventually organising a tier 2 MCP would definitely be a good idea.
Note that this is no-std no-alloc target, with all the limitations that leads to.
You could add alloc with a custom global allocator, but I don't even know what high perf global allocator you could use that wouldn't need libc. Jemalloc and mimalloc are out. Some embedded allocators would work (but those are rarely high performance, instead being optimised for small code and data footprints).
That said, with enough effort (quite a lot!) it would be possible to add support for alloc and std without libc on Linux specifically (since it has a stable syscall ABI).
What might be more realistic though is looking at relibc (a rust implementation of libc, made for Redox OS but from what I read it also supports Linux). But I haven't tried it and I don't know the state (or goal) of it.
> it would be possible to add support for alloc and std without libc on Linux specifically (since it has a stable syscall ABI).
Well yes that’s a Linux specific target so that’s kinda the point.
Technically you could do libcless on a few other platforms which are not actively hostile to it (yet) like freebsd, but that would have no chance of getting to tier 2 if it was even accepted.
Most of musl's performance issues come from their allocator. Using it with a third party high performance allocator allows you to benefit from static linking with very little performance loss.
> Most of musl's performance issues come from their allocator. Using it with a third party high performance allocator allows you to benefit from static linking with very little performance loss.
This is addressed and disputed very early in the article. The very first benchmark presented shows a 26% regression using musl + mimalloc, a high-performance 3rd party allocator.
It's not really disputed since musl without mimalloc has a 144% overhead, so most of the performance issues do indeed come from the allocator, by a pretty large margin (~85% of it). Not only that, but some of the "other code" performance hit might still come from the allocator: when you set a global allocator on the Rust side, musl still uses its own allocator internally (as demonstrated by https://github.com/BurntSushi/ripgrep/issues/3494).
And the compounding issue is that the allocator issues get significantly worse as parallelism increases, as the allocator is serial, so as concurrency increases so does the impact of the allocator, which is not the case for most of the "regular slow" code (of musl), those have a relatively constant overhead per thread.
"Using it with a third party high performance allocator allows you to benefit from static linking with very little performance loss" is disputed; 26% is not "very little," even if 144% is worse.
But the reality is that almost no one actually cares about performance because compute is cheaper than expertise and labor, at least in the short haul. Everything is getting more bloated and slower and we just compensate by adding CPU cores, gigabytes and gigahertz.
Its string routines and memory copy routines are also similarly bad, as the article alludes to. They are just naive loops with nearly no optimization. These are not small insignificant functions where using them is "doing it wrong", they are the backbone of vast amounts of code and can be made multiple times faster. You can similarly see string routines pop up in profiles all the time in musl builds in my experience. And unlike the memory allocator these cannot be "fixed" systematically across the application at link time, so you are stuck with it.
Real programs have to often do things like allocate memory and use multiple threads and process strings. People have been optimizing these things for decades, there is vast amounts of prior art, the musl developers simply did not do so because they prioritize simplicity over nearly everything else (from what I can tell) including performance.
You can buy cars and trucks that are optimized for driving on freeways and residential streets carrying stuff people often carry. But then there are special vehicles like fork lifts and such that are kinds of large special cases. And then there are weirdo specialised vehicles that have four wheels but are rare and their users can’t live without them.
Languages like C++ let you plug in special allocators if you want. But most people don’t. Some, like HFT people do crazy headstands to avoid slow allocations. I don’t ever want to do that but if they want to, why not. I don’t think they complain that the default case doesn’t fit their needs!
Doesn't have to stay that way, with hardware prices soaring and development cost allegedly in free fall.
It's wild that such a fundamental piece of code (you can't really implement operation on structs without those) is library-supplied. I wish compilers would just have something like __builtin_memcpy and __builtin_memset, and provided some highly optimized, specialist-crafted assembly in those, instead of having to inline the library code and hopefully be able to optimize it.
Glibc, for example, has perhaps ten different implementations of memcpy just for x86. The compiler certainly could provide all that, but the next step is harder:
glibc automatically dispatches to the proper one at runtime based on the actual microarchitecture that the binary is running on. You pay the extra dispatch cost once, but all of non-inline function call cost every time. This is what allows distros to compile to a nice baseline architecture, but still get near-optimal memcpy performance on many more architectures than a single inline instance could possibly give. These differences matter.
And it does it for not just memcpy, but half-a-dozen other extremely performance sensitive library functions, like strcpy and so on.
Inlining works very much against this strategy. If you can guarantee that the target microarch never changes, then it isn't a good one. But that is somewhat unusual for everyone but those who build their own binaries to run on a single class of machines forever.
Worse, inlining the really high performance versions of these ends up being terrible from a code size perspective, because they are often hundreds of instructions, which can have bad caching effects. And once you amortize the function-call cost over many iterations of the loop, it isn't so expensive to call out to the library.
Anyway, just some additional considerations to think about.
For GCC, there is -minline-all-stringops:
https://gcc.gnu.org/onlinedocs/gcc-16.2.0/gcc/x86-Options.ht...
It does what it says, but the results may not be what you expect.
The ones provided by the compilers are simply the libc ones.
LLVM will even go as far as detect attempts to rewrite memcpy and replace them with a call to the libc one!
The problems at first glance :
- Not having control over the implementation detail of the interface that your library provides is probably not wise. Sounds like a lot of bad bug reports and edge cases that you have no control over.
- Not all compilers may provide these.
Moreover the 26% is with mimalloc, with musl's allocator it's 144%, so there are likely other parts that are slower (likely the memcpy implementation)
I use musl for my Rust stuff because I have noticed that for the stuff I write it appears to have a lower memory footprint; since a lot of what I do is IO-bound anyway, I care more about using less memory than raw performance.
For a much more technical discussion, see https://github.com/sharkdp/fd/issues/710
[0]: https://news.ycombinator.com/item?id=45143347
That to me is the main driver for MUSL.
If this specific use case is of high interest to you and you have some available bandwidth, contributing to it, maybe becoming a maintainer, and eventually organising a tier 2 MCP would definitely be a good idea.
You could add alloc with a custom global allocator, but I don't even know what high perf global allocator you could use that wouldn't need libc. Jemalloc and mimalloc are out. Some embedded allocators would work (but those are rarely high performance, instead being optimised for small code and data footprints).
That said, with enough effort (quite a lot!) it would be possible to add support for alloc and std without libc on Linux specifically (since it has a stable syscall ABI).
What might be more realistic though is looking at relibc (a rust implementation of libc, made for Redox OS but from what I read it also supports Linux). But I haven't tried it and I don't know the state (or goal) of it.
Well yes that’s a Linux specific target so that’s kinda the point.
Technically you could do libcless on a few other platforms which are not actively hostile to it (yet) like freebsd, but that would have no chance of getting to tier 2 if it was even accepted.
This is addressed and disputed very early in the article. The very first benchmark presented shows a 26% regression using musl + mimalloc, a high-performance 3rd party allocator.
And the compounding issue is that the allocator issues get significantly worse as parallelism increases, as the allocator is serial, so as concurrency increases so does the impact of the allocator, which is not the case for most of the "regular slow" code (of musl), those have a relatively constant overhead per thread.