mps VS samsara

I have a library which has an extremely slow free, around 2m for large files, because of unnaturally scattered allocation patterns, but this old conservative GC didn't help at all. It was about 40% slower with libgc. mimalloc was a bit better. Best would be a properly fast GC, like mps https://github.com/Ravenbrook/mps, but this would be too much work.

The first thing you should look at is MPS (see https://github.com/Ravenbrook/mps and https://www.ravenbrook.com/project/mps/). It's open source, professionally maintained and very powerful, and it was used e.g. in Dylan and LispWorks.

I wonder how the MMTK compares to the venerable Ravenbrook MPS https://www.ravenbrook.com/project/mps/ which originated in Harlequin’s programming language implementations, particularly Dylan.

I am trying to implement garbage collection for my language because I want memory management for arrays/lists and strings. I am looking through LLVM's garbage collection page but the documentation isn't great. Are there any other resources that offer more concrete steps to implement garbage collection? Would it be wise to circumvent LLVM all together for garbage collection and only use something like the Memory Pool System? Thanks!

This post seems related to the authors of MPS (1) that seems to be a general garbage-collector to use with various languages.

Many GC'd languages really didn't bother with stack-allocating variable-size entities, and regardless of if they did then _precicely_ scanning the stack would be complicated without compiler help.

If the compiler doesn't leave any info to the GC, then it can't know if it's scanning a pointer or a float and if your GC strategy relies on compacting memory (ie moving objects) then trying to guess between a float or a pointer can become fatal.

(1) https://github.com/Ravenbrook/mps

> IME it's the other way around, per-object individual lifetimes is a rare special case

It depends on your application domain. But in most cases where objects have "individual lifetimes" you can still use reference counting, which has lower latency and memory overhead than tracing GC and interacts well with manual memory management. Tracing GC can then be "plugged in" for very specific cases, preferably using a high performance concurrent implementation much like https://github.com/chc4/samsara (for Rust) or https://github.com/pebal/sgcl (for C++).

> Just for example: "it needs a GC" could be the heart of such an argument

Rust can actually support high-performance concurrent GC, see https://github.com/chc4/samsara for an experimental implementation. But unlike other languages it gives you the option of not using it.

The compiler support you need is quite limited. Here's an implementation of cycle collection in Rust: https://github.com/chc4/samsara It's made possible because Rust can tell apart read-only and read-write references (except for interior mutable objects, but these are known to the compiler and references to them can be treated as read-write). This avoids a global stop-the-world for the entire program.

Cascading deletes are rare in practice, and if anything they are inherent to deterministic deletion, which is often a desirable property. When they're possible, one can often use arena allocation to avoid the issue altogether, since arenas are managed as a single object.

There are concurrent GC implementations for Rust, e.g. Samsara https://redvice.org/2023/samsara-garbage-collector/ https://github.com/chc4/samsara that avoid blocking, except to a minimal extent in rare cases of contention. That fits pretty well with the pattern of "doing a bit of GC every frame".

There are a number of efforts along these lines, the most interesting is probably Samsara https://github.com/chc4/samsara https://redvice.org/2023/samsara-garbage-collector/ which implements a concurrent, thread-safe GC with no global "stop the world" phase.

Nice blog post! I also wrote a concurrent reference counted cycle collector in Rust (https://github.com/chc4/samsara) though never published it to crates.io. It's neat to see the different choices that people made implementing similar goals, and dumpster works pretty differently from how I did it. I hit the same problems wrt concurrent mutation of the graph when trying to count in-degree of nodes, or adding references during a collection - I didn't even think of doing generational references and just have a RwLock...