2024836 - Experiment with preferring dirty memory even if it increases run fragmentation in chunks

Status: NEW

(Core :: Memory Allocator, enhancement)

Description

[Paul Bone [:pbone]]

Extend Bug 1898647 further to prefer dirty memory at the cost of some fragmentation. We can control how much fragmentation and how expensive the search is and should experiment with these values.

Comment 1

[Paul Bone [:pbone]]

Attachment: WIP: Bug 2024836 - Expriment with preferring larger runs if they're dirty

Try to further improve run allocation by allocating from a larger run if
that larger run is dirty and therefore won't incur a page fault.

Comment 2

[Paul Bone [:pbone]]

https://perf.compare/compare-results?baseRev=d1e4ac0cb8ae9683952459ef2cf84373fbd2f80b&baseRepo=try&newRev=abd465102772495c12ffe94a9403be50006401f9&newRepo=try&framework=13

This doesn't make much of a difference with these conservative settings (the patch includes some constants for how hard to search). Increasing those limits improves Linux performance.

Fragmentation is better for base content processes but worse for used content processes.

Comment 3

[Jens Stutte [:jstutte]]

I wonder if we could do an extreme experiment with just always taking from the largest dirty run, avoiding any search? Just to see where fragmentation would go?