Skip to content

Reference Counting Garbage Collector mode #34836

@oxinabox

Description

@oxinabox

This came up in the ML call a few weeks back.

It may be useful to have multiple different garbage collector implementations, and the ability to chose which to use at startup.
While the current Mark-and-Sweep based-GC is a generally good chose for most use-cases its not ideal for all.
A Reference Counting based-GC has a different set of trade-offs.
In particular when you have hetrogenious memory; from multiproccessing (distributed or otherwise), and especially from GPU or other accellerators, it can be desirable to have very "tight" lifespans on all memory.

Now, I am in no way saying that we can't make out Mark-and-Sweep do this well.
But I am suggesting it might be simpler to have a second garbage collector that one can manually switch to.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions