View Single Post
Old 2020-10-09, 18:23   #625
VBCurtis
 
VBCurtis's Avatar
 
"Curtis"
Feb 2005
Riverside, CA

33×5×37 Posts
Default

Quote:
Originally Posted by Mr. Odd View Post
Wow, really? Is there anything in your setup that makes it so, or do you think that's true across the board?
I hand-tuned the CADO params files for speed- they're posted in the CADO subforum https://mersenneforum.org/showthread.php?t=24274. Most of my files are 20-30% faster than the factory defaults, but the C100 file in particular is slow by default (quartic poly, for one). I think non-hyperthreaded machines will not be quite as fast running CADO as HT-enabled.

Similar work can be done for factmsieve/GGNFS, particularly on the relations-wanted values that trigger the first filtering run. I haven't run factmsieve in quite a while, but I think the stock settings filter a whole bunch of times, stalling sieving quite a bit. Edit- I've never run GGNFS controlled by Yafu, so I can't comment on its efficiency.

My best CADO C100 timing is a tick under 6 minutes wall-clock time, on a single-socket Xeon 12x2.5Ghz. On an haswell-i7 (6x3.3ghz), I have 9.5 minutes for C101 on CADO.

On C93, I have 226 sec on CADO, 229 sec on siqs (both 12-threaded on 12 cores).

EDIT: As noted in the above-referenced thread, Skylake-yafu using AVX-512 is much faster, with a crossover to CADO around 97-98 digits.

Last fiddled with by VBCurtis on 2020-10-09 at 18:34
VBCurtis is offline   Reply With Quote