This is now closer to 0.1% as of Chrome 50. The usages of this have windows of size 10, 60, and 1000. Given the microsecond precision of base::TimeDelta, duplicates are probably rare, so storing the entries in a multiset is unlikely to be beneficial.

Since it is the inserts which are slow, perhaps it would be much simpler to just use a dequeue<int64_t>, and compute percentiles by a copy, sort, index, and return a TimeDelta.