https://www.researchgate.net/publication/221321424_A_unified...
Performance characteristics depend where on that continuum your workload falls. For example, Erlang/BEAM uses a generational GC for most common heap objects, but refcounts large binary blobs. This is pretty much a perfect case for refcounting: new references are created infrequently, copying or moving is expensive, destruction is deterministic and happens immediately after the last reference disappears, and there're no pointers within the blob that would require tracing or cycle-detection.
Similarly, UI components within a GUI is another good case for refcounting (and presumably why Apple continues to use this for Objective-C and Swift in Cocoa). New references happen only in non-performance-critical code, most data remains live across collections, and copying/moving existing data would a.) be slow and b.) would invalidate any C pointers into contained data.
Sounds like the particular problem domain described in this article is one where heap allocations are frequent, which makes generational GCs more appropriate. That's probably the case with the vast majority of computational algorithms, but there are definitely problem domains where refcounting continues to beat GC.
It would leak memory - reference counting cannot collect cycles (you need tracing GC for that, defeating the purpose of refcounting).
> (though in order to be general a small amount of additional work is needed)
Not also that there are two methods of cycle detection for a reference counted GC that are not just a backup tracing-GC
1. Trial deletion (known since at least the mid 80s)
2. Various tracing systems that exploit extra information known to reference-counted systems e.g. Levanoni/Petrank[1] which actually implemented a reference counted GC for Java.
https://news.ycombinator.com/item?id=22959600
1. Reference counting is a form of GC; you could implement a JVM that used reference counting (though in order to be general a small amount of additional work is needed)
2. Reference counting causes extra work every time a reference appears or disappears. Tracing GCs amortize that cost across many allocations.
2.b. This is particularly hurtful to performance for short-lived objects, since most tracing GCs have zero GC overhead for short-lived objects (the cost of a nursery collection under most implementations scales with the amount of live data in the nursery, so objects that appear and disappear in the time-span of a single nursery GC are freed at zero extra cost). Furthermore a tracing GC
3. Malloc cannot move allocated data, so many implementations have a lot of complexity to avoid heap fragmentation, which comes at a cost to both allocating and freeing data. Many GC'd languages allocate small objects with a single instruction in the typical (just incrementing a pointer, the non-typical case would be when the nursery is full and a GC happens).
4. the JVM and Go both have a lot of effort put into their GC; the ref-counting implementation used by this test is probably a bit more naive. In particular they talk about large delays when a chain of links cause many allocations to die at the same time. A less naive refcounting implementation would queue deleted objects and spread that work out across a larger time period.