▲ 31 points
back
78 comments
> Think of a service that keeps a large cache in memory, or an index built out of millions of small objects that all point at each other. Every one of those pointers has to be followed on every cycle, for as long as the process is up.
That's a strange thing to assert, having acknowledged the existence of generational GC.
A generational GC might help with some part of the problem but does not solve it entirely.
Major collection still happen which will scan everything, just with a reduced frequency.
Long libed caches might still have pointers being mutated frequently (like an LRU cache), which will put objects in younger generation.
Golang does not have a generational GC. There are examples where long lived caches have been a problem in golang. One example - https://discord.com/blog/why-discord-is-switching-from-go-to...
Similarly, the statement
> Reference counting also has its own running cost, paid on every copy of a pointer you make and every time you drop one.
isn’t 100% true. It’s not necessarily on every copy or drop. Compilers can (and do) elide reference count updates if they can proof they aren’t needed, and can even skip allocating room for reference counts if they can proof it isn’t needed (example: a local object that doesn’t escape its scope)
I also find it a miss that the article doesn’t discuss memory usage. A garbage-collected program needs more memory to match the performance of the equivalent manually managed language.
https://dl.acm.org/doi/10.1145/1094811.1094836 says you need to give it 5 times the memory, but that’s from 2005 and likely outdated.
But you can also put generational GC on steroids by implementing escape analysis, which creates a separate category of allocations that don’t have to be marked and swept at all. Effectively they were marked the first time the code was run and never have to be marked again.
What people forget too is that new and destroy in C++ aren’t exactly cheap either. As far back as Java 5 the hotspot people were bragging on how new in Java was cheaper than new in C++, closer to stack allocation speed.
Of course if you look at any Java library other than ones intended for embedded or HFT systems, you will see Jevons’ Paradox at play. Made allocation cheap, and everyone allocates the hell out of everything, eating up the gains and then some.
But at the same time, the architectural simplicity of not having to worry about object lifetimes makes APIs simpler, which allows people to invest that brain power into making them more sophisticated in other dimensions. Watching Rust try to get itself doubly linked lists was strange to watch from the bleachers.
In 1990s C++ we did:
void foo::method(const &smart_ptr<thing> x) { ... }
It's still done. It's not uncommon to see that in Rust:
fn foo(x: &Arc<T>)
I think you can find similar things in CPython, on the C side of things.
Only if reference counting is exposed in the type system, or the standard library types are blessed to the compiler, otherwise there is no way for the compiler to know what to elide.
It occurs to me that I have never seen any description of generational GC that actually notes how that works, though.
If we have two generations, young and old, we obviously can't simply "just collect in the young generation" — which I feel like is exactly what most generational GC descriptions say. If we only traced young objects, we could very well miss an old object pointing at a young one, collect the younger object, and now we've got a dangling pointer.
So there must be some book-keeping to avoid tracing everything (or what'd be the point of generations), but I have no idea of what that book-keeping is, or its cost.
> Show me your code and conceal your data structures, and I shall continue to be mystified. Show me your data structures, and I won't usually need your code; it'll be obvious.
That's where write barriers and the like come in. The cost varies by implementation and language.
> If we only traced young objects
Careful; just because you only collect in the new generation doesn't mean that you only trace the new generation. From my understanding a rudimentary generational GC design is to include the old generation in the GC roots so you don't need to examine everything in the new generation - if it's not reachable from the roots it's implicitly dead, and the latter category is assumed to apply to most objects in the new generation under the generational hypothesis.
> just because you only collect in the new generation doesn't mean that you only trace the new generation.
Well, then I suppose you're saying the up-thread comment is wrong?
TFA:
> Every one of those pointers has to be followed on every cycle
The comment:
> That's a strange thing to assert, having acknowledged the existence of generational GC.
Which would imply that's not the case, i.e., that we're not considering every pointer in every GC sweep. (Which, again, I thought was largely the point of generational GCs: to make sweeps cheap by not considering every pointer.)
I guess if it's really the sweep that's the expensive part, then perhaps doing a full walk is fine.
> generational GC design is to include the old generation in the GC roots so you don't need to examine everything in the new generation
I'm assuming roots (stack references to objects) are separate from generations (which heap objects belong to).
I suppose if you added old objects to the set of roots, that'd also solve it, but that's the same as "every one of those pointers has to be followed on every cycle".
The sibling post thinks writes are made more expensive by tainting/young-ifying objects that get written to. In that way, we prevent an old object from ever pointing at a new one — at the cost of writes now being more than a write.
Edit: Yeah, here's [a note](https://chromium.googlesource.com/v8/v8/+/refs/heads/13.3.25...) about how Chrome implements it. There is additional book-keeping and cost to writes.
When you get it working, you will figure it out. (Or even well before then; the rest is just bugs.)
I wrote one for TXR Lisp which there are no generation heaps; but objects have a generation. In the same heap, you can have adjacent (unrelated) objects that are in different generations. Objects never move from their heap; they keep their position in the same heap over their lifetime.
The allocator therefore records the new generations in a fresh log, which serves as the nursery. When that log gets full, a pass is triggered.
Implementations with separate heaps (typically copying collectors) allocate in a nursery heap and promote objects from there, but it is conceptually similar.
In the marking phase, we do not process this nursery. We start at the usual root pointers: stack, globals and proceed with normal marking, like in a pure mark-sweep collector (or a full pass), with a small modification: whenever we hit an old gen object, we skip it. We assume that the old gen object is reachable, and every object reachable through it is similarly an old-gen object. This is where the generational algorithm wins, chopping down the graph of objects to be marked, possibly drastically so.
Where the nursery/fresh log comes in is the sweep phase. Because we know that we did not traverse any old generation objects, it would be wasteful to do a full sweep. In the case of the fresh log, we sweep just through the fresh log. Anything in the fresh log that is reachable is promoted to the old generation. Anything not reachable is reclaimed: either immediately or through the finalization treadmill, if applicable. The fresh log is then reset to empty. In the case of a nursery, we similarly just sweep through the nursery heap and promote (by copying) reachable objects to the old generation, reclaiming the rest. The nursery is empty.
There may be additional structures. Note that we have the assumption that old objects only point to old. But what if (it is allowed that) the program mutates an old object to point to a newly allocated new one? That would break the assumption. We can make the program (code generated by compiler or whatever) report whenever that happens.
In the TXR Lisp implementation of generational GC, mutation of old objects is handled via two strategies. A single value assignment of a young object to a field in an old one will cause the young object to be appended to a "check" log. In some cases, this is not practical for various reasons. For instance, an operation mutates a large number of fields of the same object (e.g. array). In such cases we add the old objects to a "mutated" log. The objects added to both arrays have their generation field reset to -1: neither young (0) nor old (1).
Both the check log and mutated log are marked during marking. The check log contains only young objects and so sweeping those is already taken care of by the freshlog, it needs not be visited during the sweep phase.
The mutated log is processed during sweep in order to reset the generations from -1 back to 1.
When the check and mutated logs fill up, GC is not triggered immediately then, but the flag is set for the next GC to be a full one. What that does is turn off the mechanism: since we know a full GC is coming, we don't have to record mutations of old objects pointing to new.
Even things built directly on underlying malloc and free typically have some form of "garbage collection" in the malloc implementation for efficiency and performance (geometric sizing, thread caching, etc).
It's best to think about lifetimes and lifecycles where possible. Immutability where sensible and things like pool allocation are examples of this.
GC languages can result in quite pessimistic code because they encourage people to NOT think about what is going on. But people have also built functional HFT engines on things like the JVM by thinking about lifetimes and lifecycles.
Ultimately, the entire problem of deallocation in general is a continuum, not a binary, and I tend to find it hard to take anyone seriously who is vigorously arguing about how awful GC is if they don't understand that and indicate some understanding of the concept. The closest you can get to real "manual" memory management on a modern system is to use nothing but arena allocations, and if you really want "manual" memory management, you need to allocate some small fixed number of arenas, because if you're constantly allocating and deallocating them that is itself probably an automated process that could go wrong, at least in theory. Malloc/free or new/delete isn't really "manual memory management".
The wide variety of options and tradeoffs, with fewer clear lines in the sand than most people seem to think, is already enough to call it a "continuum" but what really finishes the job is that they're all mixable and matchable. Something like Zig makes that really obvious, but most static languages have at least some sort of ability to mix in multiple strategies. There's nothing wrong with a C++ program that uses new & delete, and also uses arenas for some things, and also uses garbage collection for some things, and also has an integrated scripting language like Lua with its own memory strategies. Such programs are not that uncommon... that describes modern games nowadays, the supposed canonical case where you "can't afford GC". But it can... it just fences it in to a particular domain where it fits.
Careful, this statement is misleading:
> The leaks that come from forgetting to free something go away entirely.
Not quite: In manually-managed and referenced counted languages with destructors, releasing resources, (open file handle, open socket, open connection to a database, ect,) happens when objects are cleaned up.
In a (tracing) garbage collected language, releasing resources is a very manual process. You might not have a memory leak, but leaking file handles or similar resources is a real problem with real consequence.
Agreed, for most real world softwares, the cost of GC is irrelevant. But there are still some programs like databases or game engines where the cost can start adding up. That's when you measure and optimize.
Did you "Agreed" your own blog?
This is getting ridiculous.
Yet the three major game engines Unreal, Unity and Godot all have a GC on their infrastructure, and Capcom is quite happy with their .NET fork on RE Engine.
Also every single graphics application that uses Metal or DirectX, relies on reference counting as GC algorithm.
I’m sure those three engines have had no issues with performance whatsoever right?
Oh shit…
This assumes that you’re only ever running a single software at a time. Sure, 2-10x slower/memory consumption might not matter in a vacuum, but when every software is like this, you get machines that feel slower than they did 2 decades ago.
> That's when you measure and optimize.
How do you "optimize" the GC away after you wrote your entire database server in a language that uses it?
It's incredibly in common in game development, C# has a lot of features you can take advantage for this.
But the most naive example any language supports is simple object pooling.
Then more fancy, zero allocations tasks in C# https://github.com/cysharp/unitask
Object pooling and bump allocators using persistent scratch buffers, mostly. The latter what you'd use for read buffers in I/O intensive applications like databases and the like.
GC languages typically have features of the language and/or standard library that make GC the default, not the only option.
If the language is Java which puts the GC in nearly everything you're f...,
other languages have GCs and nonGC'd data so you go to the wonderfull world of manually managed objects with its use-after-free..
Starting by using a GC language that is strongly typed, compiles to native code, supports value types, memory pools/arenas if required, which provides best of both worlds.
Where GC means any kind of GC algorithm from CS point of view.
Lots of ways but an obvious and generic answer is to use pooling.
Props for using a correct nomenclature. Reference counting is a "kind of automatic garbage collection. Tracing garbage collection is also a kind of automatic garbage collection.
I used to write a lot of Javascript-like Extendscript scripts back when I was using InDesign a lot. The DOM's global object $ had a method to directly invoke the garbage collector. It made a difference certainly, but it was difficult to tell to what extent because InDesign itself gradually leaks memory and becomes more bloated the longer you use it in a single session.
There's a third way here. It could be called many things: Single ownership by default, automatic stack lifetimes, hidden unique pointers, etc. The main idea is that the lifetime of dynamic heap data is treated no differently to primitive stack data: Clean it when it goes out of scope. Rust and C++ require you to specify this manually, but Nim is unique among native-compiled languages in that is does it by default, with tools to opt-out. An advantage of this approach is that combining immutable values and static analysis can reduce most parameter passing to borrows, again, without needing the programmer to specify, by default. The main cost being that some assignments, especially crossing the variable-to-immutable line or vice-versa, would require a copy.
What does GC cost?
For Caddy with an incredibly synthetic http only benchmark it costs about 2ms of latency and somewhat less throughput.
Worth it in an incredibly artificial benchmark? Perhaps. However when it comes to real world usage the cost is a significantly smaller piece of the pie.
> In Rust you pay for it by arranging your program in a way the compiler can verify.
I disagree with this. The sentence implies that this work is done in order to make the compiler happy, where my experience is that it forces the programmer to actually get it right.
I had an "aha moment" when I was frustrated at failing to express my intent to the compiler, and suddenly realised that the reason I couldn't "just say the magic words" was that my object ownership design was inherently flawed. I had to make large changes not to make the compiler happy, but to actually have a coherent design.
So no, it's not about what "the compiler can verify". That's like saying "my lawyer won't let me do this". No, your lawyer is your employee, not your boss. They're just saying that if you do this, then you may go to prison. It's not the same thing.
("unsafe" is the Rust way to go "thank you, legal department, but I'm making a business decision to take this risk. Your concern has been noted")
I understand what you’re saying, but I read that phrase in a different way.
Let’s say you have two ways of doing the same thing: both work, both are legit and neither introduce GC bugs. The only difference between the two is that one can be verified by the compiler while the other can’t, so you are stuck with solution no. 1 although both would work.
To phrase it differently: the code that gets verified by the compiler is safe, but is all safe code verifiable by the compiler?
I’m not implying that’s the case, but that’s what I feel the author is saying.
> To phrase it differently: the code that gets verified by the compiler is safe, but is all safe code verifiable by the compiler?
Right. And this reduces to the halting problem, so in theory the compiler cannot know that all safe code is safe.
In practice, I'm saying that not just syntactically, but in your code's design, the compiler is more likely to be right. It's a bit like Chesterton's fence. You can bypass the lifetime checks if you just have the confidence to say "yes, I'll use `unsafe` here and it's fine because these reasons". As you're writing your "SAFETY" comment, you may very well find yourself not so confident anymore. And indeed, often this compiler-induced "stop and think" prevented you steaming ahead with a bug.
Now, the borrow checker is not perfect. I don't know how far away from "all but NP-complete cases" it is. My experience is that it's almost always right, and I've only had to put a seemingly needless "drop" statement to placate it. But they're working on it. A new one is coming: https://daily.dev/posts/rust-s-new-borrow-checker-is-coming-...
And once again this old blog post of mine comes to mind: https://blog.habets.se/2020/12/Bypassing-safety-check-for-ob...
In any case "by arranging your program in a way the compiler can verify" I think is not accurate, because the overlap between "correct" and "compiler can verify" is nearly complete, though yes the latter is a strict subset of the former. In other words I don't write Rust to make the compiler be able to verify it, but to make it correct. And nearly always that means the compiler can verify it too.
I've had that moment in Rust too, where there was no composition that effected what I wanted, because supporting that model would have meant a very different backing data structure. However in that case I actually didn't care about the kind of correctness problem it was saving me from.
I've also had the converse experience, where I know full well that the structure I'm trying to impose is correct and quite efficient, but its part of the space that rust doesn't cover.
Rust is great. It's a noble attempt to bring a degree of correctness to a problem space that suffers from a great deal of slop. But to pretend that the model is complete, or that the design decisions that were made are perfect in every way, is just wrong. That the rust compiler and runtime can't support my construct isn't really an absolute value judgement on that idea in the first place. The rust compiler isn't really an oracle that tells you whether something is right or not in an arbitrary value system.
Right. The borrow checker is not always right. Just almost always right. It's not perfect (because halting problem), but… well I already said the rest in https://news.ycombinator.com/item?id=49287460
Still, I don't write Rust code the way I do "to make the compiler happy", but to make it correct. And sometimes it's correct to drop some "unsafe" because gosh darn it, you know it's fine this time.
And you're probably right for the cases you're thinking of, where Rust wouldn't let you (at least without unsafe). And maybe you're 99% sure about that.
But for every 100 changes we’re 99% sure won’t cause an outage, one will…
(also future changes may invalidate assumptions you relied on, of course, making it no longer true)
Do you have some examples you can share where you think Rust prevents you doing the right thing? The ones I run into tend to force me to think of the edge cases, and usually those edge cases don't even have a right answer.
> I disagree with this. The sentence implies that this work is done in order to make the compiler happy, where my experience is that it forces the programmer to actually get it right.
Haskellers say the exact same thing. :P Personally, I'd much rather get shit done and don't appreciate tools "forcing" me.
Which is why the software world is full of technical debt and frustrating buggy software.
But I agree that it's must faster to get the wrong answer than the right one.
I rarely see a real use case bottle necked on garbage collection.
This is one of the best high-level survey explanations of GC I have seen, nice work.
Not really the point of the post, but there is a third memory allocation paradigm where you can do manual allocations and deallocations, but without the mental overhead of having to remember to free memory because the compiler will force you to properly handle it. Mojo has linear type support so you can do something like this: https://x.com/melodyogonna/status/2085089269484343725?s=20
There is no need for garbage collection if you don’t form circular references. Just have a canonical direction and always keep weak references the other way.
Reference counting is a different model. Many papers have explored the differences and similarities, and your comment leaves so much out that it cannot even be said to be true or false.
I'd say GC is always the fastest to free objects within the main code path. Literally zero instructions.
It doesn't leave anything relevant out.
Java pioneered this garbage collection stuff because you had cycles of references. You don't need to have cycles. WeakRef is a much better thing now. All you need is reference counting, and you don't need any garbage collection at all. When the reference count reaches 0, you destroy the object and free up its memory. It's far more predictable than GC, too.
And GC isn't "the fastest" to free objects, it has to walk a graph. The fastest is actually arena allocation and then just dropping the whole thing. But that's exactly what owning an entire container of objects can do. If you have a doubly linked list, for example, A[n] -> A[n+1] but also A[n+1] -> A[n] but neither of those should be a strong reference to prevent reclaiming. Instead, the container of that doubly linked list should be the one having a strong reference to its items.
> I'd say GC is always the fastest to free objects within the main code path. Literally zero instructions
True, but reference counting or free need not be far behind. They can append the pointer being freed to a per-thread list (⇒ no locking needed) that a separate thread that does the actual freeing periodically claims and then iterates over to actually free the objects.
Disadvantage is that memory usage goes up a bit because the actual freeing is delayed, but that (likely) is less so than with a garbage collector.
GC, zero instructions: that's funny; JVM used to "stop to world" to process that zero instructions;