It's hard to reason about how much noise is guaranteed to be enough, because it depends on how much measurement the adversary has a chance to do, there could be collusion beween several sites, etc. To allow timing API usage I'd be more inclined toward the virtual time thing I mentioned upthread.