Tweaking polynomial search for C197 - Page 4

bai · 2011-06-15, 03:40

Quote:

Originally Posted by jrk

Note that the lambda parameters differ between ggnfs lasieve4 and CADO. In ggnfs, the base for lambda is the factor base bound while in CADO it is the large prime bound. So you should use larger lambda with ggnfs, probably about 2.6 for rlambda and 3.6 for alambda in this case. Else you will lose a lot of relations.

Thanks! I see.

With 2.6 and 3.6, for c5 = 124620, I got

Quote:

total yield: 53, q=300000119 (1.67887 sec/rel)

For the other polynomial, with 2.6 and 3.6, it gives

Quote:

total yield: 70, q=300000119 (1.62300 sec/rel)

jrk · 2011-06-20, 02:38

Quote:

Originally Posted by bai

Could you please try to use I=15 for the sieving test (would be better than 16 for this size I guess?).

Might be stretching a bit to use 15 for sieving a c197.

Quote:

Originally Posted by bai

Thanks! I see.

With 2.6 and 3.6, for c5 = 124620, I got

Quote:

total yield: 53, q=300000119 (1.67887 sec/rel)

For the other polynomial, with 2.6 and 3.6, it gives

Quote:

total yield: 70, q=300000119 (1.62300 sec/rel)

The sample size seems too small. For GNFS sieving especially, with so few specialq lattices sieved for each poly, there's going to be a lot of noise in the results. We usually sieve a few thousand specialq lattices for each poly to make a comparison, like fivemack detailed earlier. This would take several CPU hours (or days) sieving for each poly, but would be much more precise.

bai · 2011-06-23, 09:20

Quote:

Originally Posted by jrk

The sample size seems too small. For GNFS sieving especially, with so few specialq lattices sieved for each poly, there's going to be a lot of noise in the results.

Yep, definitely my test sieving is too short.

Batalov · 2011-07-01, 02:21

After a brief cerias.purdue.edu downtime, a new line showed up:
6054 2,1196+ c197 74501884005720267921233705046612300434042100618110806127512805905183882417313. p120 Bos+Kleinjung gnfs

But Tom's c197 is a tad larger. Anyway, when data for B+K factorization will show up, the strength of the polyns could be compared.

jrk · 2011-07-07, 20:20

Quote:

Originally Posted by fivemack

Please download the most recent SVN, add

Code:

	{195, 1.00E+030, 1.00E+029, 2.00E-015},
	{197, 1.00E+030, 1.00E+029, 2.00E-015},

to the end of the definition of prebuilt_params_deg5 in gnfs/poly/poly_skew.c

and recompile. It should then work.

It'd be nice if we could finish up the deg5 param table in msieve. Do you think the above params are adequate for this size job now? Or have you made changes to them yet?

Beyond this, params for 200-215 digits would be just a guess right now (deg6 starts at about 220 digits).

fivemack · 2011-07-08, 08:12

My conclusion was that, for jobs at the 180-digit and larger scale, you need one more parameter than that set of options allows: I got much better results by running -np2 separately and with a much lower middle parameter than was optimal to get reasonable numbers of hits at -np1.

But I haven't worked that up into an actual patch.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
c176 polynomial search	bdodson	Msieve	45	2010-10-29 19:39
109!+1 polynomial search	fivemack	Factoring	122	2009-02-24 07:03
5^421-1 polynomial search	fivemack	Factoring	61	2008-07-21 11:16
6^383+1 by GNFS (polynomial search; now complete)	fivemack	Factoring	20	2007-12-26 10:36
GNFS polynomial search tools	JHansen	Factoring	0	2004-11-07 12:15

2011-07-01, 02:21	#37
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 3⁶×13 Posts	After a brief cerias.purdue.edu downtime, a new line showed up: 6054 2,1196+ c197 74501884005720267921233705046612300434042100618110806127512805905183882417313. p120 Bos+Kleinjung gnfs But Tom's c197 is a tad larger. Anyway, when data for B+K factorization will show up, the strength of the polyns could be compared.

2011-07-08, 08:12	#39
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 7²·131 Posts	My conclusion was that, for jobs at the 180-digit and larger scale, you need one more parameter than that set of options allows: I got much better results by running -np2 separately and with a much lower middle parameter than was optimal to get reasonable numbers of hits at -np1. But I haven't worked that up into an actual patch.