swcl VS gc-latency-experiment

related: the Immix inspired parallel-mark-region GC developed by Hayley Patton (https://github.com/no-defun-allowed/swcl) got merged recently into SBCL.

https://github.com/sbcl/sbcl/blob/master/doc/internals-notes...

https://applied-langua.ge/~hayley/swcl-gc.pdf

build with

    ./make.sh --without-gencgc --with-mark-region-gc (on x86-64/Linux and x86-64/macOS only at the moment).

The Immix inspired mark-region GC developed by Hayley Patton (https://github.com/no-defun-allowed/swcl) got merged recently, which is pretty cool news for SBCL users.

Slides about the ongoing parallel garbage collector for SBCL: https://applied-langua.ge/~hayley/swcl-presentation.pdf (https://github.com/no-defun-allowed/swcl)

There is this document which details how the collector works.

You want some steel wool for that, or...

I'm currently implementing Immix in SBCL (got generations working yesterday), but I wouldn't recommend one does that for a miniature implementation, as my implementation (though with hacks, performance tricks of dubious quality, and having to put the hysterical raisins somewhere) is about 1kLOC, and I haven't even gotten to writing compaction yet. A simple hybrid of mark-sweep and copying is the Lang-Dupont collector which copies one portion of the heap at a time, marking and reclaiming the rest in place.

I say this while failing to design a parallel GC for SBCL for my second time; but I was unaware of the more clever RC schemes until recently, as nor myself nor my older colleagues seemed to have only used naive RC.

"the distributed system" is the first link, garbage collector is this link; again I wrote the backend for SICL as well as other parts, and I didn't say I made a language.

There's this repo which is more for latency than throughput, and the annoying part of a message bus. But I think it's too cruel on copying collectors, as it only creates large-ish byte arrays that a collector has to keep copying around; moving from a generational-copying collector to a generational-mark-region collector cut latency down about a factor 4 on its own.