but saying they're useless ignores a bunch of real systems that wouldn't run without them.
in unity, you literally can't do burst compiled jobs efficiently unless you choose the right allocator. they expose `Temp`, `Persistent`, etc because GC isn't even an option when you're fighting for milliseconds per frame. no allocator choice = frame skips = shipped bugs.
in embedded, FreeRTOS gives you multiple heap implementations for a reason. sometimes you need fixed size pools, sometimes you care about fragmentation. malloc's out of the question. same in any real time firmware, safety critical or not.
infra world has been using arena allocators for years. folly's `Arena`, grpc's arena, even jemalloc has thread caching and slab style region reuse built in. this isn't academic. large scale systems hit allocation pressure that general purpose allocators can't tune for globally.
and rust's whole alloc abstraction : it's more about expressing ownership across lifetimes in memory sensitive codebases. `Bumpalo` isn't a premature optimization. it's what you reach for when you know the object graph layout and want to free it in one call. also safe by design. it's not even painful to use.
imo allocator choice is an interface decision. it shapes how the rest of your code handles memory lifetimes. once you start passing memory lifetimes as part of your type system or job model, the whole architecture shifts. this is way deeper than faster malloc
I have an STM32 rust project I've been working on this week. It talks to an ESP using protobuf/RPC.
I'm doing it bare-metal/no allocator, as I do most embedded projects... and it's flirting with running me out of memory! What do most (and the most popular) protobuf libs do in rust? Use an allocator. What does the ESP itself do? Use an allocator (with FreeRTOS).
Meanwhile I'm using Heapless (Vec and String syntax with a statically-allocated array), on a MCU with 128K flash and 32K Ram... This won't end well.
On first seeing this I wasn't sure what analogy the author was trying to make. After reading the article my best guess is that they are simply trying to say that, writing an allocator is easier than it seems on the surface.
Though it's not clear to me that the article does a good job of establishing that this is actually true ("mimalloc is only a few thousand lines of code" doesn't pass the smell test).
Yeah, re: the title, the URL/path string "allocators-are-for-monkeys-with-typewriters" (including on the page) seems more clear about that "less hard than you'd think" thing than the larger-font published headline "Allocators are Monkeys With Typewriters". And of course the quote in the article is even more specific "given enough time, even a monkey with a typewriter can write a memory allocator".
I generally agree with the "memory management doesn't have to be as complicated as you might think" vibe, especially if you've read about some optimizations in fancy modern GCs and aren't aware of what a basic simple non-GC world can look like. That said, of course, you can indeed get into a lot of complexity beyond the textbook 101 examples. Like the mentioned threading...
Applications should use more special-purpose memory allocators. Much of the complexity in memory management is designing for an unknown usage pattern and things can be quite simple when allocator is specialized and patterns are predictable.
This is difficult though in higher-level languages. Go tried and failed with arenas.
in unity, you literally can't do burst compiled jobs efficiently unless you choose the right allocator. they expose `Temp`, `Persistent`, etc because GC isn't even an option when you're fighting for milliseconds per frame. no allocator choice = frame skips = shipped bugs.
in embedded, FreeRTOS gives you multiple heap implementations for a reason. sometimes you need fixed size pools, sometimes you care about fragmentation. malloc's out of the question. same in any real time firmware, safety critical or not.
infra world has been using arena allocators for years. folly's `Arena`, grpc's arena, even jemalloc has thread caching and slab style region reuse built in. this isn't academic. large scale systems hit allocation pressure that general purpose allocators can't tune for globally.
and rust's whole alloc abstraction : it's more about expressing ownership across lifetimes in memory sensitive codebases. `Bumpalo` isn't a premature optimization. it's what you reach for when you know the object graph layout and want to free it in one call. also safe by design. it's not even painful to use.
imo allocator choice is an interface decision. it shapes how the rest of your code handles memory lifetimes. once you start passing memory lifetimes as part of your type system or job model, the whole architecture shifts. this is way deeper than faster malloc
I'm doing it bare-metal/no allocator, as I do most embedded projects... and it's flirting with running me out of memory! What do most (and the most popular) protobuf libs do in rust? Use an allocator. What does the ESP itself do? Use an allocator (with FreeRTOS).
Meanwhile I'm using Heapless (Vec and String syntax with a statically-allocated array), on a MCU with 128K flash and 32K Ram... This won't end well.
Though it's not clear to me that the article does a good job of establishing that this is actually true ("mimalloc is only a few thousand lines of code" doesn't pass the smell test).
I generally agree with the "memory management doesn't have to be as complicated as you might think" vibe, especially if you've read about some optimizations in fancy modern GCs and aren't aware of what a basic simple non-GC world can look like. That said, of course, you can indeed get into a lot of complexity beyond the textbook 101 examples. Like the mentioned threading...
This is difficult though in higher-level languages. Go tried and failed with arenas.
but it's just more blah-blah about ai :/
https://www.youtube.com/watch?v=LIb3L4vKZ7U
...I'll be here all week. Try the veal!