3 ms·

compaction really does help runtimes alot. but I'm not sure how much it really has to do with line level locality. in general we don't try to batch related obje

compaction really does help runtimes alot. but I'm not sure how much it really has to do with line level locality. in general we don't try to batch related objects together except in a coarse form by generation. I think the measurable benefit comes from page level savings, both reducing the number of trips to the kernel to get zeroes pages, and from reduced pressure on the tlb. but I have definitely seen numbers like 20% on some workloads for turning on compaction

tsimionescu 3y ago
> in general we don't try to batch related objects together except in a coarse form by generation. It greatly depends on the GC algorithm, right? Copying collectors naturally bunch up related objects (though of course, in which order is less well defined, if one object has multiple pointers to others).