If you're going to LLR test the entire file, then one should sieve until the factorremoval rate is half as fast (twice the seconds per factor) as an LLR test takes for the smallest candidate in the file.
This is due to the way the sieve scales when you take candidates out of the sieve for LLR, the sieve speed does not change linearly. Since there is not much speed to be gained by taking candidates out, we leave the small candidates in longer than commonsense might suggest.
A rough scaling example: If you take out 10% of the candidates from the sieve file, the sieve only runs ~5% faster ("faster" as measured by p/sec).
