Fix constraint detection in the GC #51656
Conversation
oscardssmith
added
performance
| if (constrained_mem > 0 && constrained_mem < total_mem) | ||
| uint64_t constrained_mem = uv_get_constrained_memory(); // This returns != 0 for not constrained | ||
| int constrained = constrained_mem != 0 ? 1 : 0; // But it can return -1 if there is a constraint but no value | ||
| constrained_mem = (total_mem < constrained) ? total_mem : constrained_mem; // The constrained value can also be greater so take the min |
There was a problem hiding this comment.
| constrained_mem = (total_mem < constrained) ? total_mem : constrained_mem; // The constrained value can also be greater so take the min | |
| constrained_mem = constrained_mem * 4 / 3; // n.b. the user constraint might be larger than the `uv_get_total_memory` value. Try to fill no more than 75% of that constraint with GC pages. |
There was a problem hiding this comment.
I don't know that this is exactly a "fix", since the behavior before was not wrong, but possibly also not intended. However, I don't think either is correct. The constraint is just the cgroup, it doesn't make sense to me to arbitrarily take the minimum with a different value and create an unnecessary constraint with that, which was neither present before nor requested by the user.
|
So my goal is to detect if we have a constraint, and if so respect it. Though if we have a constraint that doesn't have a set value (like GHA does) we probably want to follow the max physical memory at least (GHA does set it). Though we might want to make a bit fancier and do something like. If it's uintmax, then follow max physical, otherwise follow the constraint set? |
|
We should probably respect |
to the constraint instead of a fixed amount.
|
About the |
Sure. make a PR? |
gbaraldi
removed
the
backport 1.10
This wasn't respecting the documentation that Libuv has, which lead to weird results. This might have caused some issues with julia inside containers