The object copy prediction during concurrent mark is a bit too conservative, it assumes object copying is slower during concurrent mark (which it shouldn't be). Investigate if this is the case. If it isn't, then rewrite the policy to always use the same predictions for object copy.