GH-139951: Fix major GC performance regression #140262
Conversation
* Count number of actually tracked objects, instead of trackable objects. This ensures that untracking tuples has the desired effect of reducing GC overhead * Do not track most untrackable tuples during creation. This prevents large numbers of small tuples causing execessive GCs.
markshannon
added
needs backport to 3.14
|
I'm not sure I get what android (x86_64) fails.
|
|
I'm not at all familiar with the garbage collector, but one of the ways that Android differs from the other platforms is that it runs all the test suite serially in a single process. |
| } | ||
| intptr_t new_objects = gcstate->young.count; | ||
| intptr_t max_heap_fraction = new_objects*3/2; | ||
| intptr_t max_heap_fraction = new_objects*2; |
There was a problem hiding this comment.
The comment above is not in line with this one, but I don't want to stop this PR from merge now. We can fix this in follow up PR.
|
Should we redo the benchmarks to see how the final version impacts performance? Maybe we can run it on https://github.com/faster-cpython/benchmarking-public (but I don't know how). |
| PyThreadState *tstate = _PyThreadState_GET(); | ||
| struct _gc_runtime_state *gcstate = &tstate->interp->gc; |
There was a problem hiding this comment.
| PyThreadState *tstate = _PyThreadState_GET(); | |
| struct _gc_runtime_state *gcstate = &tstate->interp->gc; | |
| struct _gc_runtime_state *gcstate = &_PyInterpreterState_GET()->gc; |
Maybe it'll be better to use the same expression as below?
There's no another call of _PyThreadState_GET in this file.
| small tuples. | ||
|
|
||
| * Counts number of actually tracked objects, instead of trackable objects. | ||
| This ensures that untracking tuples has the desired effect of reducing GC overhead |
There was a problem hiding this comment.
| This ensures that untracking tuples has the desired effect of reducing GC overhead | |
| This ensures that untracking tuples has the desired effect of reducing GC overhead. |
| /* Fast, but conservative check if an object maybe tracked | ||
| May return true for an object that is not tracked, |
There was a problem hiding this comment.
| /* Fast, but conservative check if an object maybe tracked | |
| May return true for an object that is not tracked, | |
| /* Fast, but conservative check if an object maybe tracked. | |
| May return true for an object that is not tracked. |
5 participants
This PR:
For the example in the original report this makes performance on main a bit better than 3.13.
Benchmarking results show this is about neutral on performance otherwise.