pol51 going nuts - Page 2

fivemack · 2009-04-01, 22:09

I was looking for normmax parameters for very large jobs:

At 165 digits, 10^24.8 / 10^23.3 / 1.4e-16 seems reasonable

At 170 digits, 10^25.7 / 10^24.2 / 5.0e-17 seems reasonable

At 175 digits, 10^26.2 / 10^25.0 / 1.0e-17

At 180, 10^27.5 / 10^25.8 / 2.3e-18 seems reasonable; normmax1 of 10^26.5 or 10^27.0 generates no first-round hits.

(definition of 'seems reasonable' is that the normmax2 is reached by about 10% of the things that pass the normmax1, and the murphy score is reached by about 1% of the things that pass the normmax2)

I think in an ideal world you'd adjust normmax1 by looking at the number of outputs per second from pol51m0b, but you need a good model of how fast pol51opt is to figure out what the right rate to aim at would be.

Batalov · 2009-04-01, 23:29

Excellent. I'll dig up some small jobs. I had tarred them from last summer, I probably had a dozen or more.

The feeling is that the trendlines (which are frozen in def-nm-params.txt) are slanted a bit wrongly; so while it all works in the middle, it doesn't do as well on both ends:
-- no yield for high-end; I remember that I've trial-end-error'd good normmax's for the GNFS-156 and GNFS-160 and they were far from simply interpolating the top of the table,
-- ...and an avalanche at low end which breaks the levees in the pol51m0b.

So with your high-end tests and easy low-end tests, the better fit could be simply planted into the perl script. I'll take on it over the weekend.

fivemack · 2009-04-02, 06:59

Umm, note that I did those high-end tests with msieve rather than with the pol5 suite; whilst the two programs are computing the same norms, there's no guarantee that the asymptotics are the same!

jasonp · 2009-04-02, 14:05

Serge, do you remember the 160-digit polyselect parameters you used? Adding them in would mean a smooth progression of parameters above 155 digits...

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Der Geoibils now require logs+nuts(to be pronounced with New Joisey accent)	c10ck3r	Lounge	8	2013-11-18 22:57
bowl of nuts	Fusion_power	Puzzles	11	2012-09-22 18:49
Question about pol51	jrk	Factoring	1	2010-01-19 04:02

2009-04-01, 22:09	#12
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 7²·131 Posts	I was looking for normmax parameters for very large jobs: At 165 digits, 10^24.8 / 10^23.3 / 1.4e-16 seems reasonable At 170 digits, 10^25.7 / 10^24.2 / 5.0e-17 seems reasonable At 175 digits, 10^26.2 / 10^25.0 / 1.0e-17 At 180, 10^27.5 / 10^25.8 / 2.3e-18 seems reasonable; normmax1 of 10^26.5 or 10^27.0 generates no first-round hits. (definition of 'seems reasonable' is that the normmax2 is reached by about 10% of the things that pass the normmax1, and the murphy score is reached by about 1% of the things that pass the normmax2) I think in an ideal world you'd adjust normmax1 by looking at the number of outputs per second from pol51m0b, but you need a good model of how fast pol51opt is to figure out what the right rate to aim at would be.

2009-04-01, 23:29	#13
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 9477₁₀ Posts	Excellent. I'll dig up some small jobs. I had tarred them from last summer, I probably had a dozen or more. The feeling is that the trendlines (which are frozen in def-nm-params.txt) are slanted a bit wrongly; so while it all works in the middle, it doesn't do as well on both ends: -- no yield for high-end; I remember that I've trial-end-error'd good normmax's for the GNFS-156 and GNFS-160 and they were far from simply interpolating the top of the table, -- ...and an avalanche at low end which breaks the levees in the pol51m0b. So with your high-end tests and easy low-end tests, the better fit could be simply planted into the perl script. I'll take on it over the weekend.

2009-04-02, 06:59	#14
fivemack (loop (#_fork)) Feb 2006 Cambridge, England 7²·131 Posts	Umm, note that I did those high-end tests with msieve rather than with the pol5 suite; whilst the two programs are computing the same norms, there's no guarantee that the asymptotics are the same!

2009-04-02, 14:05	#15
jasonp Tribal Bullet Oct 2004 3,541 Posts	Serge, do you remember the 160-digit polyselect parameters you used? Adding them in would mean a smooth progression of parameters above 155 digits...