[jblow]

It is simply not true unless you have a pathological idea of "hand-allocation" (which to be fair, some programmers do program like).

Let me put it this way ... all "garbage collection is fast" claims are saying the following thing:

"It is faster for the programmer to destroy information about his program's memory use (by not putting that information into the program), and to have the runtime system dynamically rediscover that information via a constantly-running global search and then use what it gleans to somehow be fast, than it is for the programmer to just exploit the information that he already knows."

It sure sounds like nonsense to me.

[_ph_]

You are talking about an abstract case. Yes, if an objects lifetime is perfectly easy to determine, manual memory management is faster. But in many typical programs with real dynamic memory requirements, this is not the case. One has to keep track about the life time of objects in the program logic. Either by storing additional information as for reference counting, or careful designing your program logic so the status of an object is always clearly known.

The latter might not bear a cost in CPU cycles, but certainly in programmer cycles, if the program logic allows for it at all. It also puts the burden of correctness on the programmer vs the garbage collector. And for most contemporary programs, correctness is the larger challenge then execution speed, especially with GCs which can run in parallel with your program logic, so can utilize unused CPUs "for free".

[mcguire]

"This paper presents a tracing and simulation-based experimental methodology that executes unaltered Java programs as if they used explicit memory management. We use this framework to compare the time-space performance of a range of garbage collectors to ex- plicit memory management with the Lea memory allocator. Com- paring runtime, space consumption, and virtual memory footprints over a range of benchmarks, we show that the runtime performance of the best-performing garbage collector is competitive with ex- plicit memory management when given enough memory. In par- ticular, when garbage collection has five times as much memory as required, its runtime performance matches or slightly exceeds that of explicit memory management. However, garbage collec- tion’s performance degrades substantially when it must use smaller heaps. With three times as much memory, it runs 17% slower on average, and with twice as much memory, it runs 70% slower. Gar- bage collection also is more susceptible to paging when physical memory is scarce. In such conditions, all of the garbage collectors we examine here suffer order-of-magnitude performance penalties relative to explicit memory management."

http://people.cs.umass.edu/~emery/pubs/gcvsmalloc.pdf [2005]

Consider what happens when you allocate or free memory manually: the manager had to do some amount of work to track the status of the memory.

Now consider a 2-space, moving collector. Allocation is a pointer bump and range check. Collection means tracing and copying the live objects. That's it.

It's not as simple as it seems.

[vvanders]

Yup, an arena allocator that you drop at the end of an operation will always be faster than a GC.

[yosefk]

Well, yeah, but if you need some of the allocated data at the end of the operation, you'll end up copying it, which will have a runtime cost, might introduce bugs (because of pointer invalidation and the ease of forgetting to update some pointers) and then none of the two will get counted as "cost of manual memory management." C++ with its value semantics encourages unnecessary copying tremendously and it's never counted as "cost of memory allocation" or "producing garbage", on the contrary, Stroustrup says things like "C++ is my favorite GC language because it generates so little garbage to begin with."

This is not to say that it's fair to call OCaml "efficient" in the memory department based on a GC benchmark; TFA is full of examples where OCaml allocates things on the heap that TFA recommends to allocate elsewhere and shows you how to maul your code to get there.

My only point is that how well a programming system uses memory is a very hard question because (A) there are many different use cases and (B) you can't isolate "memory performance" into a few easily measurable things like time spent allocating, time spent in GC and peak memory use - there are other things like what your program has to do outside of the allocator to cope with its semantics and how the performance of code using the memory objects is affected by the layout encouraged by the allocator and these things cannot be measured in isolation from the rest of the program.

[pjmlp]

Not if you have cascaded data structures, have some kind of destructor like operations to call, synchronize threads, or if the memory allocator does a call into the OS to release the memory back to the OS.

[mcguire]

Two facts of computer science: Destructors suck. And destructors are wonderful.

[pjmlp]

I loved Herb Sutter's talk at CppCon 2016, where he showed how careless written destructors can lead to "stop-the-world" and stack overflows in complex data structures.

[mcguire]

Arenas are probably the fastest memory management technique this side of fixed allocation and stacks.

If the lifetime of the objects is contained within the lifetime of the arena. If not, you have to copy live objects out before you destroy the arena, at which point you are implementing a n-space moving collector.

[rbehrends]

For a modern generational GC, the minor heap works just like an arena allocator (at least for the data that can be thrown away), so this shouldn't result in any performance difference (assuming the minor heap is large enough).

[gsb]

I think that when you find arguments that GC can be competitive with manual memory management, they always rely on the assumption that it is in the context of idiomatic code. For a GC language, idiomatic code allows the use of a generational collector which can automatically operate in a manner similar to an arena allocator. I suppose this is a matter of opinion, but I think that using arena allocators in a language like c++ is not strictly idiomatic - you should only add them when you find they are required for optimisation, and they have a cognitive overhead. If you make the wrong choice with your arena allocator then either you can make performance worse or you can introduce memory safety problems.

So this is the way that a GC might be competitive with manual memory management - if the benefit of arena-like allocation offsets the additional tracing performed by the GC.

[_yosefk]

Isn't it harder to deallocate the data if you actually check that it's unused, as a GC must, instead of just assuming it?

[barrkel]

A GC doesn't check if memory is unused; it looks for memory that's used and frees whatever is leftover. Having a small fraction of used memory to total memory is what makes GC cheap (in compute time). It also means that allocating short-lived memory with abandon is cheap, if that's most of your allocation.

[_yosefk]

Got it, an interesting point and it follows that a use case can be engineered where a GC adds essentially zero overhead over an arena allocator.

But in the general case where some objects are short-lived and others aren't, surely manually splitting your allocations to malloc for long-lived ones and some sort of arena_alloc for short-lived ones ought to be faster than allocating them all in one place, then copying the ones which are still reachable out of the area reserved for short-lived objects?

(This is not to say a GC-based system will be slower "on average" because nobody knows what the "average" is. A realistic arena-based system can have objects most of which are short-lived but some do need to live longer and you only find out long after they're allocated; in that case, one has to manually reallocate those objects just like a GC would, and doing it, say, the C++ way is definitely more bug-prone than GC's bug-free handling of this, and one way to make it less bug-prone in C++ is to have deeper copies and avoid trying to minimize copying, and now you might easily be slower than a GC. I'm just saying that it's very easy to find a case when a system not getting any hints from the programmer wrt object lifecycles and instead discovering them fully automatically would be slower than a system which does get these hints. And of course a GC can provide ways to supply these hints, I'm just not aware of one which does - perhaps it's avoided on the theory that the GC algorithm might change and you don't want to make hints which operate in terms not portable between algorithms a part of your interface.)

[rbehrends]

> But in the general case where some objects are short-lived and others aren't, surely manually splitting your allocations to malloc for long-lived ones and some sort of arena_alloc for short-lived ones ought to be faster than allocating them all in one place, then copying the ones which are still reachable out of the area reserved for short-lived objects?

This is where things get difficult, actually, and you need benchmarks because:

1. You still have a bump allocator that's much faster than a general purpose first-fit/best-fit allocator and has generally better memory locality than a pool allocator.

2. Offsetting that may be the additional tracing and/or copying you are doing as a result of garbage collection.

3. Manual memory management techniques often have their own performance costs: std::shared_ptr and std::unique_ptr both add overhead and you sometimes see additional copying where lifetimes are difficult to predict.

Which one has the higher cost is often something that can be only tested for with actual code (and can go either way).

I'll also note that this is primarily of interest for functional languages, which often have a high allocation rate. For imperative programs, the memory allocation rate (and ratio of memory that contains pointers to memory that doesn't) is often so low that even a very basic mark/sweep collector would be fine, as long as pause times don't matter in your application domain. This means that it's often not really a practical issue, one way or the other.

[barrkel]

IMO the design of the memory system has a strong influence (or should have a strong influence) on the architecture of a system when performance is a big concern. What the "average" is, is under the control of the program itself; a program written for a generational GC should be written differently to one with a refcount-based GC, and both should be different to one for a simple mark and sweep GC.

Similarly, designing a program with manual allocation for speed means using some kind of zone or arena allocation, probably mixed with some stack-oriented allocation that mirrors the control stack (not necessarily consuming CPU stack space). The design of the memory system needs to be integrated with the architecture of the program and the lifecycle of the values it needs to track.

I don't think there's any simple winner. Different problem spaces require different treatments of memory. For example, stateless servers have little need for long-lived memory; they're well suited to generational GCs, but also to arenas (though GC is easier to keep correct, and usually more fluent in practice, without rewriting too much of the standard library). Mix in in-process caching, and things start getting murkier. Put caching in a different process, keeping it simple, at the cost of some IPC; tradeoffs, etc.

[gsg]

In general you can't drop the contents of the arena on the floor in a GCed language, since you might have been wrong about those objects being short-lived (or because they are short lived, but they were allocated shortly before the young heap filled up and happen to still be alive). So tracing/copying is still necessary.

It is quite possible for objects that are expected to tenure to be allocated directly into the major heap. Many systems use various criteria to decide when this should be done. In particular it is common to allocate very large objects directly to avoid ever having to copy them.

There are also static analyses which attempt to determine which objects can be safely allocated in an arena-like way, that is, region inference. In some systems in addition to inference the programmer is given ways to specify that a given object should live within a certain region. MLkit is the usual example of such a system. Region systems don't seem to have been particularly popular, but there's still some work being done on them here and there.

[rbehrends]

Unused data is by definition unreachable from any GC roots and should not incur additional tracing work.

[ori_b]

That means that the heap grows with the highest address allocated within it, unless you incur additional overhead for copying the data around.

[rbehrends]

This is exactly what a generational GC does. Once the minor heap fills up, it evicts data that is still in use and moves it to the major heap.

Note that this overhead is only incurred for data still in use. We are not comparing memory management strategies in general, but the special case of temporarily allocating data that can be thrown away at the end.

[gsb]

Absolutely, but I think the "in most cases" of the original quote was meant to imply simple malloc/free style memory management rather than the use of tailored arena/region allocators.

[rwmj]

That could blow away your cache if the arena is large. Better to measure it.

[_lbaq]

Heap allocation and smartpointers (as an alternative to GC) is by no means free, in that context GC performance favorable in many scenarios, for example when doing large number of small allocations.

But in the end, if your application have serious performance requirements, you should avoid allocating memory in the hot path altogether.

[_ph_]

So much this. GCs don't mean you shouldn't care about reasonable memory allocations. In hot paths, not allocating will always be faster than allocating. Having a GC does not prevent you from using object pools and buffers, where it is a clear performance win. But having a GC means, you are not required to do so, where there is no bottle neck. The difference between a "fast" and a "slow" GCed language often is, how well the programmer keeps some control about memory allocations.

Fun fact: the Go GC is written in pure Go code, which does not perform any heap allocations. So there is enough control to write code in Go which does not require any heap allocations. There is even a compiler switch for the development of the GC code which makes heap allocations a compile error.