FTA: “The implementation of Green Tea has a special case for pages that have only a single object to scan. This helps reduce regressions, but doesn’t completely eliminate them.”

Also FTA: “One surprise result of this work was that scanning a mere 2% of a page at a time can yield improvements over the graph flood.”

⇒ I think you’d have to try and get two objects on each page, and they would have to be small (you’d have to be able to fit over 100 objects in a page to have 2 live objects be <2% of all objects in the page)