I'll take a ~1day go at this one with CADO.

I haven't found anything better than the best score I posted after searching the full recommended A-range (up to 40M+). One interesting thing to note is that by the time all was said and done, the polynomial that produced that top score isn't even in the best 400 from the -nps step.

The best e-score came from a polynomial that produced a Stage 2 norm of 1.930819e+021. When the full sieving was finished, limiting to the best 400 norms cut off at 1.719606e+021.

I think I'm going to run -npr on the full ms file, just to see if anything shows up.

[QUOTE=wombatman;433922]
The best e-score came from a polynomial that produced a Stage 2 norm of 1.930819e+021. When the full sieving was finished, limiting to the best 400 norms cut off at 1.719606e+021.[/QUOTE]

This is quite interesting! I've been doing poly select by aiming for 100-150 hits per day from -nps, and -npr'ing once a day on that file. So, a 5- day run might generate 500-700 hits, and I wondered if I was wasting npr effort. I'll stop wondering!

I normally do a fairly arbitrary reduction of the Stage 1 norm by 20 and the Stage 2 norm by 10. Other than that, I collect everything output by the -nps step, sort, and then typically have grabbed the best few hundred (by lowest norm) to do -npr on. It seems like it may be beneficial to either do the -npr more frequently as you describe, or to do -npr on a larger section of the -nps results (or all).

Here's the twenty best lognorms CADO produced (ETA on rootopt whenever I wake up, probably ~12 hours from this post):

[code]# Stat: best logmu after size optimization: 47.66 47.69 48.09 48.10 48.28 48.37 48.39 48.45 48.55 48.58 48.60 48.63 48.73 48.74 48.74 48.78 48.79 48.82 48.83 48.83[/code]

For comparison, e^48 ~ 7e20. My 400th best hit has lognorm 49.85 ~ norm 4.46e21, and my 900th best and cutoff (900 being the CADO recommended quantity to rootopt) is lognorm 50.19 ~ norm 6.27e21.

IIRC msieve and CADO norms are directly comparable, so it seems that GPU-Msieve (I assume that's what y'all are using?) does beat CADO on sizeopt on sheer brute force power.

Edit: Within the first few dozen rootopted polys I have a best hit of 1.05e-12. I believe it's going in order of best norm, so we'll see what pops up.

Here's the top five polys from my overall CADO run:

[code]n: 429029423121896674206591410060656286271312034858546242336901864226665117427625764458140583950338797912894203455691819016559421597936003179046137489332768988628331
skew: 706304.0
c0: 3010021158731930833855111739761917440
c1: 6335561415094488563952374760588
c2: -129589712732642425420134442
c3: -20538593223066435741
c4: 92030654876030
c5: 26586300
Y0: -9579492148677276851109960722289
Y1: 552853772915427493477
# MurphyE (Bf=1.00e+07,Bg=5.00e+06,area=1.00e+16) = 1.09e-12
# lognorm 50.53

n: 429029423121896674206591410060656286271312034858546242336901864226665117427625764458140583950338797912894203455691819016559421597936003179046137489332768988628331
skew: 706816.0
c0: 1909563021037302710148241915210422960
c1: 492980532041025556792738716838
c2: -16737408643641489200622009
c3: -6566991973742394623
c4: 4296764452964
c5: 9423120
Y0: -9813722336863454034628330883771
Y1: 7618374674620938497387
# MurphyE (Bf=1.00e+07,Bg=5.00e+06,area=1.00e+16) = 1.05e-12
# lognorm 49.01

n: 429029423121896674206591410060656286271312034858546242336901864226665117427625764458140583950338797912894203455691819016559421597936003179046137489332768988628331
skew: 740096.0
c0: 1004340719161714036902664542098409936
c1: 11230382949364358280843740453230
c2: -2330599688710933689459049
c3: -32392368087958757597
c4: 17849498811090
c5: 7350840
Y0: -12847581150086001832828718268543
Y1: 2371008964221663336371
# MurphyE (Bf=1.00e+07,Bg=5.00e+06,area=1.00e+16) = 1e-12
# lognorm 49.15

n: 429029423121896674206591410060656286271312034858546242336901864226665117427625764458140583950338797912894203455691819016559421597936003179046137489332768988628331
skew: 370304.0
c0: 195175623906042759954065926176202971
c1: 4078845487257653550917186346955
c2: -13105083218979618302495721
c3: -24261316345406015767
c4: 63574955077482
c5: 50040720
Y0: -7619727016390169240329141063972
Y1: 8125184165455301559871
# MurphyE (Bf=1.00e+07,Bg=5.00e+06,area=1.00e+16) = 9.74e-13
# lognorm 49.24

n: 429029423121896674206591410060656286271312034858546242336901864226665117427625764458140583950338797912894203455691819016559421597936003179046137489332768988628331
skew: 615168.0
c0: 5326200533922631393675172974957866600
c1: 1227651854691868107862805625522
c2: -47643043732458779413660997
c3: 19506919860043479037
c4: 37864118088088
c5: 46097280
Y0: -9427131594106150989448594510066
Y1: 103505366839884201599
# MurphyE (Bf=1.00e+07,Bg=5.00e+06,area=1.00e+16) = 9.64e-13
# lognorm 50.41
[/code]

GW_4_424
 

The following C165 needs a polynomial.
[CODE]105621823495048537976278940353371952215164776400133723221602262267919915507914311657609017352042009951185865123618356040681037053727141536311042495817323967136828767[/CODE]

On it, starting at A5=1. Will run through the full A-range (max is ~52.5M or so) before doing -npr. GPU portion will probably take a day or two.

[QUOTE=wombatman;433962]I normally do a fairly arbitrary reduction of the Stage 1 norm by 20 and the Stage 2 norm by 10. Other than that, I collect everything output by the -nps step, sort, and then typically have grabbed the best few hundred (by lowest norm) to do -npr on. It seems like it may be beneficial to either do the -npr more frequently as you describe, or to do -npr on a larger section of the -nps results (or all).[/QUOTE]

I think if you reverse the reductions, you'll obtain better results. I divide stage 1 by 9 or 10, stage 2 by at least 20 (I've been using 25 to 30 recently, but obv 20 will yield more hits). Reducing stage 1 by 20 is giving up hit-rate in -nps. Try two runs with the same input number and same stage 2 norm. One with stage 1 reduced by 10, the other reduced by 20. I'm pretty certain the former will produce ~30-50% more hits in the .ms file.

Interesting. I'll try that out.

Clarification: I meant hits per unit time, rather than hits per A-range. It's obvious that a higher stage1 norm will produce more hits per A, but not very relevant. I run -np1 -nps together, and ran a bunch of trials to see what settings maximized the number of quality hits per day.