> It is far more predictable than GC but I don't believe that it isn't more expensive.

A GC has to determine liveness of objects, so it has to traverse the object graph (or at least the most recent generation if the GC is generational) to know which objects are still alive and which aren't. That's what "GC scaling" refers to. Then it still has to deallocate dead objects, and pay the same deallocation cost as scope-bound free.

Good generational GC doesn't need to deallocate objects one by one, instead of that it can deallocate objects in bulk (if there is no finalizers). It can allocate memory faster because it doesn't need to use memory allocator to create every object, it just increment pointer that points to the beginning of the free space. It's true that modern GC has greater throughput than simple allocation/deallocation schemes. Problem is predictability (as been mentioned).

It doesn't have to deallocate dead objects. That is sort of the point of most collectors.

All GCs have to deallocate objects. It may be very cheap to do so, for example when you have a semispace collector and no finalizers, but switching the spaces is still a form of deallocation. Production-quality GCs don't use semispace for everything because of the huge memory use overhead, and all practical GC'd languages I know of need finalizers in some form.

Finalizers are certainly the exception rather than the rule for objects. Deallocation to me represents some action taken for a particular object and scales with the number of objects deallocated. Semi-space copying collectors do not have this property. I think we only disagree on the semantics of "deallocation" — which is why I specified my argument in terms of how things scale. Similarly compacting collectors like Java's CMS also have this scaling property without semi-spaces.

But it has to scan/mark all the objects, not just the the unused ones. This has large overhead in big-mem apps. Not to mention that usually this forces the whole app into RAM (makes swap useless) and also potentialy drains cpu caches. And to amortize this it often has more memory overhead (up to 2x) than typical fragmentation in a manually managed app.

I'm a little bit confused. Every object that gets allocated has to be freed at some point. That cost is the same no matter how the memory is managed.

Are you talking about an additional cost beyond "freeing memory that was allocated"?

It sounds like they are talking about a semi-space collector, which I can imagine might in theory be faster than scoped allocation since allocation is a simple pointer bump (just like stack allocation) but when garbage collecting you only have to traverse the live set. However, semi-space collectors can't collect non-memory resources without additional metadata, and they have a memory overhead of at least 2x the maximum live set. I have a hard time imagining even a semispace collector outperforming stack allocation.

Of course if your program is short-lived you can just allocate a sufficiently large heap, pointer bump the whole way and free all non-memory resources on program termination, which ought to be cheaper (but probably not that much in practice due to cache effects).

In Rust, you often use stack allocation, which, as you pointed out, is much more efficient. There's some heap allocation in Rust, but in my experience it's actually a minority of all allocations.

Rust also has an Arena library (http://doc.rust-lang.org/0.11.0/arena/index.html), which can be used for cases where you know that you have enough memory for a given operation and are willing to throw out the entire thing when the operation (or program ;) ) is done.

That link is outdated, see http://doc.rust-lang.org/arena/.

Arena allocation has similar efficiencies as GC.

Except for not having the mark phase (also not having to copy). The mark phase can take a lot of time.

Not necessarily the same. For example for arena allocation, you can drop lots of related objects at the same time. If you don't have to run destructors on them, the cost will be much lower than any other way of managing them.

I use and love Rust, but I must admit this is true. GC has some advantages over the current options. Fear not however as post 1.0, optional GC will be most likely added to Rust. See https://github.com/rust-lang/rfcs/pull/244.

In general, idiomatic Rust puts most objects that would be in a semispace nursery on the stack instead, which has all the advantages of a semispace collector, plus improved locality and the lack of a mark phase. The lifetime system means that the traditional limitations of escape analysis (higher order functions, separately compiled functions, virtual functions, returning objects on the stack) largely don't apply.

If you can get all the objects for a full request on the stack, that would be pretty great. Is that true for the current HTTP servers / frameworks written in Rust?

I don't do a lot of server Web development in Rust, but I see no reason why it wouldn't be possible, at least for common cases with small vector optimization.