Currently G1 does not try to make sure that threads on a given NUMA node preferentially process objects/work from the same node.

They just take whatever work item/object they can get and then copy it to the correct location.