mersenneforum.org Call for volunteers: RSA896
 Register FAQ Search Today's Posts Mark Forums Read

 2013-02-03, 18:04 #78 Stargate38     "Daniel Jackson" May 2011 14285714285714285714 17·37 Posts How many hits are needed? What's the ETA after the required number of hits is reached?
 2013-02-03, 19:14 #79 jasonp Tribal Bullet     Oct 2004 2·3·19·31 Posts The answer to your first question is that nobody knows. On theoretical grounds we're supposed to be able to find polynomials that are drastically better than what we've been seeing, so if the 'time constant' of the search is measured by the time it takes to get really lucky once, then we've barely started. The previous search for RSA768 generated 30 million hits, one of which turned into a polynomial whose performance was indistinguishable from the one actually used in 2008-2010 My archives now contain 21 million hits (thanks everybody!), and Paul's group has posted 18 million more found with CADO NFS (they have a large number of CPUs searching). All of the hits we've found have passed through the CADO size optimization (except Greg's large batch above), and though we've found some polynomials in their top 10 we haven't gotten near their current best. That being said, only a little of the current dataset (perhaps 5%) has passed through Msieve's size optimization. Could I persuade everyone to take a break and run the size optimization on the 39 million hits we have? This uses CPU only, and the current code can optimize a single polynomial in about 0.5 seconds, so we're looking at about 30 CPU-weeks of work to isolate the ~5000 best-scoring polynomials. Running the root sieve on those will take about 5 CPU days. To avoid me having to upload a blizzard of files, can I get 8 CPUs worth of volunteers? This job will be much easier if you can use unix text-processing tools, and I'll tell you how to run Msieve and postprocess the output. (Paul's group estimated that their best polynomial as of a few months ago would require 35000 CPU-years to complete the sieving) Last fiddled with by jasonp on 2013-02-03 at 22:56 Reason: corrected crappy time estimate
 2013-02-03, 19:33 #80 debrouxl     Sep 2009 977 Posts I can use 4 hyperthreads of the usual Xeon E3-1230 through MPI-patched root optimization.
 2013-02-03, 20:16 #81 Dubslow Basketry That Evening!     "Bunslow the Bold" Jun 2011 40
2013-02-03, 21:15   #82
WraithX

Mar 2006

1D816 Posts

Quote:
 Originally Posted by jasonp To avoid me having to upload a blizzard of files, can I get 8 CPUs worth of volunteers? This job will be much easier if you can use unix text-processing tools, and I'll tell you how to run Msieve and postprocess the output.
I have a 12-core (dual 6-core Xeon 5645) linux computer that I can dedicate to this. I'd be happy to help out.

 2013-02-03, 22:16 #83 jasonp Tribal Bullet     Oct 2004 67168 Posts Okay, I'll start uploading files. Instructions: - download your file (they're ~130MB) - unzip and rename to msieve.dat.m - with worktodo.ini set for RSA896, run Code: msieve -v -i -nps "stage2_norm=1e100" - the resulting output file is msieve.dat.ms; find the best size scores with Code: sort -g -k 11 msieve.dat.ms | head -5000 > sizeopt.out (Note for windows that this is the unix sort, not the crappy MS sort) If you want to manually split the file to run across multiple cores, rename each piece to .m and add '-s ' to the msieve command line. The optimized output will be in .ms - send me your top hits somehow. If you can't stand the suspense, rename your list of best hits to .ms and run the root sieve yourself with Code: msieve -v -i -npr -s  Note that it isn't necessary to run the root sieve on all 5000 candidates; the odds are overwhelming that the best result your batch produces will come from one of the top 100 size-optimized polynomials. Also, note that the latest Msieve SVN has an adjustment to the alpha computation that will bias the E-value downwards, but the modified score exactly matches what the CADO tools report. I'll update this post with the other files (8 total). Each file has 5M hits, and would take about 30 days on one core. debrouxl: download 1 WraithX: download 2 dubslow: download 3 firejuggler: download 4 WraithX: download 5 dubslow: download 6 poily: download 7 debrouxl: download 8 WraithX: Greg's pile in post #110 Last fiddled with by jasonp on 2013-02-09 at 18:16
 2013-02-03, 23:19 #84 Dubslow Basketry That Evening!     "Bunslow the Bold" Jun 2011 40
 2013-02-03, 23:46 #85 firejuggler     Apr 2010 Over the rainbow 5·503 Posts i'll claim the third file if nobody else want it
 2013-02-03, 23:52 #86 Dubslow Basketry That Evening!     "Bunslow the Bold" Jun 2011 40
2013-02-04, 04:35   #87
WraithX

Mar 2006

23·59 Posts

Quote:
I'll take number 5 too.

 2013-02-04, 06:13 #88 Dubslow Basketry That Evening!     "Bunslow the Bold" Jun 2011 40 | head -<2>, where <1> is the beginning line number, and <2> is how many lines. So to split up a 1000 line file into 4 chunks, you'd do Code: for num in 1 251 501 751; do echo $num more +$num | head -250 > .\$num done` (or something similar). Last fiddled with by Dubslow on 2013-02-04 at 07:02 Reason: fix script

 Similar Threads Thread Thread Starter Forum Replies Last Post axn GPU Computing 28 2012-05-28 12:05 jasonp Operation Kibibit 200 2011-11-05 21:31 Wacky NFSNET Discussion 13 2005-07-14 00:25 Xyzzy Hardware 23 2003-04-18 23:27 Xyzzy PrimeNet 8 2003-02-27 02:26

All times are UTC. The time now is 06:17.

Sun Jan 17 06:17:45 UTC 2021 up 45 days, 2:29, 0 users, load averages: 1.67, 1.64, 1.57