View Single Post
Old 2020-03-27, 18:45   #110
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

101638 Posts
Default

Using CADO's ETA as metric, we're 65% faster with A=31 than A=32. ETA for 600M relations (CADO default, I didn't set it) is now 12 months rather than 20.

Using the 'top' CPU time versus # of relations found in a workunit, the host machine on A=32 was finding relations at 6.04 thread-sec/rel; it is running 30-threaded on a 20-core machine, so let's call that 4 cpusec/rel.

I'll update this post in a bit with the 'top' metric on A=31.
Update: With A=31, I get yield around 3.2 as opposed to 3.9 on A=32, but 2.83 threadsec/rel rather than 6.0. A=32 was not a good idea for this job! On the bright side, we got lots of relations in a Q-range that Greg doesn't search, so we helped the factorization quite a lot.
Definitely *don't* set override A=32!

Last fiddled with by VBCurtis on 2020-03-27 at 18:55
VBCurtis is offline   Reply With Quote