[rbehrends]

1. I think you may not realize what state of the art tracing GCs can accomplish. IBM's Metronome has pause times down to hundreds of microseconds.

2. It only takes freeing a tree with a few thousand nodes for it to become an issue. It happens in C++, too (heck, there've been cases where chained destructor calls overflowed the stack [1]). The reason why you don't hear more about it is because pause times just aren't that big a deal for most applications. In forum debates, people always discuss triple A video games and OS kernels and such, but in practice, only a minority of programmers actually have to deal with something even approaching hard real time requirements. Generally, most applications optimize more for throughput rather than pause times.

3. Yes, and it can be a problem for C/C++, too. It's rare, but not non-existent. Note that pools can actually make fragmentation worse for long-running processes.

4. Weak references work if you get them right. But for long-running processes, even a single error can accumulate over time.

> On the GC side, it seems like you typically get bursty, unpredictable performance, in both time and memory. Modern GCs work very hard to keep collection pauses as short as possible, but almost inevitably that means keeping garbage around for longer, which means using a lot of memory.

This ... is not at all how garbage collectors work, especially where real time is concerned. Not even remotely. I recommend "The Garbage Collection Handbook" (the 2011 edition) for a better overview. And ultra-low pause times are generally more of an opt-in feature, because they're rarely needed.

[1] E.g. Herb Sutter's talk at C++Con 2016: https://www.youtube.com/watch?v=JfmTagWcqoE&t=16m23s

[iainmerrick]

> > almost inevitably that means keeping garbage around for longer, which means using a lot of memory.

> This ... is not at all how garbage collectors work, especially where real time is concerned.

Hmm, I'm certainly no GC expert, but is it really not the case that GC tends to be memory-hungry? Not exotic academic systems, but the languages people use day-to-day.

Most of my experience with GCs is in languages like Java and C#. Java in particular can be very fast but always seems to be memory-hungry, using like 4x the memory you'd need in C++. I haven't spent a huge amount of time fine-tuning the GC settings (it seems like Oracle is working to simplify that -- good!) but the defaults seem to assume at least 2x memory usage as elbow room for the GC.

That's on the server. On mobile, I've worked with iOS and Android and iOS undeniably gets the same work done with much less memory. Flagship Android phone have 4GB of memory and need it, whereas Apple hasn't felt the need to bump up memory so quickly even after going 64-bit across the board.

The last I heard about real-time GC, with guaranteed space and time bounds, it sounded like it was theoretically solved, but not used much in practice because it was too slow. That was a number of years ago though. Has that situation changed? Are there prominent languages or systems with real-time GC?

[iainmerrick]

Looking up IBM's Metronome led me to the Jikes RVM (https://en.wikipedia.org/wiki/Jikes_RVM), which sounds so cool that I wonder why it isn't being used everywhere?

The PowerPC (or ppc) and IA-32 (or Intel x86, 32-bit) instruction set architectures are supported by Jikes RVM.

Ah, no ARM and no x64, that'd be it.

What's keeping this kind of GC technology back from the mainstream?

[rbehrends]

> Jikes RVM

The Jikes RVM is designed for research, not production. It's pretty impressive, but (inter alia) does not implement all of Java and does not support as many platforms.

> What's keeping this kind of GC technology back from the mainstream?

The fact that successful commercialization is possible; the GC tech that you see in Metronome and C4 is seriously non-trivial and not easy to reproduce unless you spend money on it; and it's also technology that businesses are willing to pay for.

At the same time, only a minority of open source use cases really require this kind of hard real-time GC, so there's little pressure to create an open source equivalent. Shenandoah is the one open source GC that does try to compete in this space, and it is trading away some performance for getting ultra-low pause times.

I'll add that this is difficult only because of concurrency and arbitrary sharing of data between threads. If you have one heap per thread, then it becomes much, much easier (and is a solved problem if soft real-time is all you need).

[throwaway91111]

One note, in video games allocations are a major source of slowdown; don't allocate in your inner loop! Use object pools and arena allocators.

[rbehrends]

This is because naive can-do-it-all allocations in C/C++ can be expensive, not because allocations are inherently expensive. In C/C++, you have:

1. A call of a library function that typically cannot be inline.

2. Analysis of the object size in order to pick the right pool or a more general allocator to allocate from.

3. A traditional malloc() implementation needs to also use a global lock; thread-local allocators are comparatively rare.

4. For large objects, a complex first-fit/best-fit algorithm with potentially high complexity has to be used.

Modern GCs typically use a bump allocator, which is an arena allocator in all but name. In OCaml or on the JVM, an allocation is a pointer increment and comparison.

Even without bump allocators, it's easy for a GC implementation to automatically turn most allocations into pool allocations that can be inlined.

Also: much as people love to talk about video games, video games with such strict performance requirements are not only just a part of the video game industry, they are a tiny part of the software industry.

[iainmerrick]

In OCaml or on the JVM, an allocation is a pointer increment and comparison.

That's true, but if (hopefully rarely) the object turns out to be needed later, it has to be copied to another heap, and that takes time and memory. Pointers need to be redirected and that takes a little work too.

Bump allocators are definitely a huge win, as good as anything you can do in C/C++ and much more convenient for the programmer, but they're not a completely free lunch.