Anything more dynamic will require some “ad hoc, informally-specified, bug-ridden, slow implementation of a GC”.
Some think that reference counting gives better performance predictability than a tracing GC, but that is no longer true, certainly compared to the generational ZGC presented here. If you could determine the time where the collection would take place with a refcounting GC, i.e. when the last reference is cleared, you wouldn't need a refcounting GC. It is unpredictable pretty much by definition.
There's a great paper (linked from this post https://www.cs.cornell.edu/courses/cs6120/2019fa/blog/unifie...) showing how refcounting GCs can match the performance of tracing GCs -- in fact, how they can each have the properties of the other -- but that requires sophistication in the implementation of refcounting GCs that most of them don't have. So it is true that a sophisticated refcounting GC and a sophisticated tracing GC can exhibit similar behaviour, but while OpenJDK's tracing GCs are very sophisticated (especially ZGC and G1), most refcounting GCs in the wild are quite unsophisticated. It may certainly be true that unsophisticated refcounting GCs have more desirable properties than unsophisticated tracing GCs, but in practice the choice offered is between unsophisticated refcounting GCs and very sophisticated tracing GCs. The latter beat the former on most counts except for footprint overhead.
I don't think I've mentioned performance anywhere; thanks for the paper I'll take a look.
I don't think so. You don't know when memory will be freed or how long that would take only that it will be freed as soon as possible. The end result is lower footprint but that's about it.
That's an unsubstantiated claim.
While reference counting may be slower than tracing in the general case, Rust's reference counting is not the same thing as general reference counting in languages like e.g. Swift or Python. There are two reasons:
- The good majority (99%+) of objects in Rust don't need to be managed by reference counting at all.
- Rust relies heavily on move semantics to avoid updating the ref-counts, so the good majority of refcounted objects have very few (sometimes just 1) refcount updates.
> reference counting gives better performance predictability than a tracing GC, but that is no longer true, certainly compared to the generational ZGC presented here.
Where does ZGC claim nanosecond latency? But even if it did, it would be still worse, because refcounting doesn't do pauses at all. Pausing applies to pausing all application threads. Spending X nanoseconds freeing memory by one thread does not count as a pause, because the system can still make progress.
> refcounting doesn't do pauses at all
At the cost of throughput.
BTW, I'm not comparing Java to Rust, just ZGC to Rust's GC. I am well aware that it is not used for most objects.
And BTW, I wouldn't be so sure about throughput either. ZGC is capable of allocation rates of the order of only ~5 GB/s, which is way below the theoretical memory bandwidth. I bet jemalloc can do much better.
> ZGC is capable of allocation rates of the order of only ~5 GB/s
That was the non-generational ZGC. Indeed, ZGC's max throughput was well below that of ParallelGC and G1. One of the main goals of generational ZGC is to increase the throughput, and as you can see in the video, they got a 4x increase in throughput.
Object layout is the primary reason for the performance differences between low-level languages and manages ones — Java uses a thread local allocation buffer, which is for all practical purposes just as hot as the stack. What is more expensive is arrays of references and such, but that will be solved by value types once, hopefully.
EDIT: What I’m trying to say is that there is very little overhead to creating short-lived objects in Java.
I've been working on high performance Java software (databases) for many years now and the overhead of creating short lived objects is bearable only because we're going so far trying not to create those objects, writing Java in a kinda C-style - no objects, just primitives as much as you can.
The total overhead is not just bumping up the pointer and zeroing memory but actually all what happens when the GC runs out of the slab. By allocating short lived objects youre making GC collect earlier and overall doing more work. Also the faster you allocate, the more objects survive the collection and more objects must be copied. That process is very expensive and also very complex.