The approach I used was to take a snapshot, count up all the different object types (making a hash table mapping "name of type of object" => "count of objects whose type has that name", then discarding the snapshot), then take another snapshot a few minutes later, and see what the difference was, and which type of object there was suddenly a lot more of. Then I looked at various instances of that object. It turned out to be some async queue-related object that was only used in a couple of places, so that narrowed it down a lot. Even if it were something generic like a hash table or list, I suspect looking at instances of the object and breaking them down by some observable quality (e.g. number of elements, the set of keys in a table, the types of objects in a list), plus the differential approach, will take you fairly far.