From the linked article. Can anyone confirm or deny this? I have hard time believing this statement.
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.
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.
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.
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.)
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.
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.
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.
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.
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.
But in the end, if your application have serious performance requirements, you should avoid allocating memory in the hot path altogether.
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.
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".
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.
- no value types mean, that there are tons of separate heap allocations which have to be referenced by pointers, this puts a ton of unnecessary load on the GC. It also means, that you fully depend on escape analysis when passing objects to functions for avoiding heap allocation. And the existing primitive types (int, float) often require boxing.
- the design of the standard libraries was, especially at the beginning full of wasteful allocations. Most string handling code was a constant series of allocations (having 16 byte chars didn't help this either).
- far to many Java libraries build extremely complex object hierarchies and protocols. Just reading from a text file means instantiating several objects. This adds to the GC pressure.
> The OCaml garbage collector is a modern hybrid generational/incremental collector which outperforms hand-allocation in most cases.
Ocaml is mostly functional and likes to allocate many short lived objects. With enough memory, a moving collector is very good at handling that load.
> Unlike the Java GC, which gives GCs a bad name, the OCaml GC doesn't allocate huge amounts of memory at start-up, nor does it appear to have arbitrary fixed limits that need to be overridden by hand.
Java's GC (-s; there's a bunch) use a heap with a fixed maximum size and an initial allocation. It has also received much more research attention over the decades. It also has a lot of knobs to fiddle with.
I've run production Java web app servers with multi-gig heaps where almost all requests were handled in the young generation. The knobs and visibility were very nice.
Ocaml doesn't use a fixed size heap, so it can conceivably take over all of memory. It also doesn't have all of the knobs. But it works pretty well.
The second problem I have with the paper is that it underestimates the kinds of structures which GC makes possible -- eg sets of trees sharing nodes. And the corollary to that is that to implement those structures with malloc/free, programmers tend to reach for ref-counting, which adds to memory usage and is generally a terrible form of garbage collection for many other reasons.
Nevertheless it's an interesting paper which does add to the debate (I studied it about 10 years ago). It is worth reading, even though I have concerns about the methodology and hence the results.
Let's say you are correct, that the manual memory management presented in the paper is too unrealistic. What do you think real world manual memory management would bring the multiple to? 5x? 4x?
> The second problem I have with the paper is that it underestimates the kinds of structures which GC makes possible -- eg sets of trees sharing nodes. And the corollary to that is that to implement those structures with malloc/free, programmers tend to reach for ref-counting, which adds to memory usage and is generally a terrible form of garbage collection for many other reasons.
These data structures are extremely niche, and when used when sharing is not truly needed just lead to time/space inefficiencies.