One more thought on that.
My realization of delays, as a single huge table, can work if we need delays which are different, but not so different. If a chunk of the table where needs to be read is fit into cache, this will should give the optimal performance.
In other words, there may be a few tables for short, medium, and long delays.
It seems the same appraoch as in lookup-tables implementation (code is here) can be used.