Repository navigation
Mark all objects reachable from roots as live before doing main cyclic GC pass #126491
Description
Activity
Is that benchmark public, link isn't working for me?
Fixed the link
FYI, I made an incremental "mark alive" version of my idea. My PR is based off of 3.13 but it could be done on top of the "incremental gc" that's in the "main" branch as well. This approach avoids a problem with the current incremental GC: if you happen to pull in an object near the root of the object graph, the GC pause could be long because all reachable objects are pull in to the cyclic GC pass.
- added a commit that references this issue
on Nov 18, 2024 - added a commit that references this issue
on Dec 2, 2024 @markshannon Out of curiosity:
- Which kind of objects are not reachable by the transitive closure of the global roots + stack, but are reachable in some other way so they are not collected by the GC?
- Are there any statistics on the ratio of number of objects (or number of cycles) that iskept by the GC to the number that is deallocated by the GC?
Python doesn't have a precise set of roots. For example, an extension module can create objects but not have them connected to any roots that the CPython runtime knows about.
I've reverted the last change, which introduced a refleak.
- added 3 commits that reference this issue
on Jan 8, 2025 Could this be closed now? Or, are we waiting for a revised version of gh-127519 to be merged? @markshannon
Objects can only be cyclic garbage if they are not reachable.
So, if we can cheaply identify the majority of reachable objects before performing the (relatively slow) cycle detecting pass, we can save a lot of time.
Performing a transitive closure of all objects reachable from global roots (the sys and builtins modules as well as builtin class's dicts and sublasses) plus a transitive closure of all objects reachable from the stacks can eliminate >90% of all objects relatively cheaply.
Initial experiments show a ~3% speedup, with an almost 50% speedup of the most gc-heavy benchmark.
This idea has been proposed a few times.
@nascheme has definitely suggested it before. Perhaps he can add links to any prior discussion and/or experiments?
What makes this more feasible now is that the GC can see the evaluation stack of frames, thanks to #124392, so we would now expect that the vast majority of reachable objects can be cheaply marked, thus improving the efficiency of cycle detection considerably.
Linked PRs