Implementing switch to disable generational behaviour of the GC

[gbaraldi]

While experimenting with julia as a shared lib the GC can become a latency issue. Specifically full GCs can cause significant pauses. While taking a look it dawned on me that some workflows don't allocate any amount of permanent memory and all of it is short lived, so turning any of it old is wasteful anyway.
I'm proposing adding a switch that turns off full gcs and promoting objects to the old age. Python has something similar called freeze https://docs.python.org/3/library/gc.html#gc.freeze

[d-netto]

I'm proposing adding a switch that turns off full gcs and promoting objects to the old age.

So you will have a single generation of objects? If that's the case, doesn't this make literally every GC a full GC? Am I misunderstanding something?

[d-netto]

Ah, OK. You may be proposing doing this after some objects have been allocated in the heap: i.e. accumulate a bunch of objects in the old generation, and from that point on, don't move anything to the old generation.

But since you had promotions before the generational behavior was disabled, then you may possibly have some objects in the old generation.

I.e. you are not really disabling generational GC at startup? (Which would make all GCs full GCs as I mentioned above).

[oscardssmith]

This isn't something you would run at startup. This is something you run after you have allocated all the memory your process needs that will live for ~the length of the process, and the rest of your process is will only have relatively short lived allocations.

[gbaraldi]

It's a switch you enable during the process. Basically you do something like

setup()
GC.gc(true)
GC.gc(true) # make sure anything that's alive is now old
GC.generational(false)
...

The idea is to make the performance more predictable.