Quote:
Originally Posted by VBCurtis
The number is a percentage of total relations, so 0.05 asks CADO to make sure there are 5% more relations than strictly necessary for filtering...

This is a decent first approximation to what required_excess does in practice, but strictly this isn't quite correct. It's not a proportion of total relations, but rather the number of relations after the first singleton removal step, which must be at least 5% more than the number of ideals remaining at that stage. This will probably require slightly more than 5% additional raw relations.
Edit: just saw bur's post. I'm seeing about 700k threadseconds per 1M range, or about 0.1% of what I estimate the sieving time ought to be on the same hardware. So 10M is still a "low" leading coefficient in this search and I'd stick with incr=420. If the returns above 10M aren't good then we can consider switching to larger incr.