Queue management for 16e queue

[VBCurtis]

jyb-
Greg, the NFS@home host, has requested our 16e jobs maintain lim's that average 225M or smaller. Is it ok to edit your job to change lim's from 268M to 225M?

[jyb]

Originally Posted by VBCurtis

jyb-
Greg, the NFS@home host, has requested our 16e jobs maintain lim's that average 225M or smaller. Is it ok to edit your job to change lim's from 268M to 225M?

Sure. Can we leave the other parameters as is, and just assume that we’ll need a little extra sieving?

[VBCurtis]

Yep! Yield will barely drop, I think; a 20% change in lim's isn't much.
Note for future jobs: mfb 96 with LP 32 is rarely best. usually 96 is used for 33, with 93-94 used for 32. Usually, a bit of yield is lost but sec/rel improves. The idea is that splitting a 96 bit cofactor is very unlikely to generate 32-32-32 split, so most of that effort is wasted.

[jyb]

Originally Posted by VBCurtis

Yep! Yield will barely drop, I think; a 20% change in lim's isn't much.
Note for future jobs: mfb 96 with LP 32 is rarely best. usually 96 is used for 33, with 93-94 used for 32. Usually, a bit of yield is lost but sec/rel improves. The idea is that splitting a 96 bit cofactor is very unlikely to generate 32-32-32 split, so most of that effort is wasted.

Good to know, thanks.

[swellman]

C235_131_106 from the XYYXF project. A strange one to be sure.

I test sieved both the deg 5 and deg 6 SNFS polys. Neither sieves all that efficiently, I assume due to the awkward coefficients. But the quintic clearly is more efficient in sieving - not even close. Maybe the sextic coefficients are just that much more awkward?

After numerous test runs, the best way I could find to submit this to NFS@Home is as a 16f/31-bit job. Felt a bit counterintuitive, but what do I know.

15e struggled to get any yield on either poly, though I did not try 15e/33-bit. Seemed a bit much. 16f/31 was the sweet spot.

Any objections to me submitting this job?

n: 1030503235456803762016167471530641995037760274197968310988179590505950995510892101498055733828425469497568921009021217030573300085480516986964919661166160134908753817168257662257684691416593544651963249784543552924184296481653156578163
# 131^106+106^131, difficulty: 267.34, anorm: 2.36e+032, rnorm: 4.45e+058
# scaled difficulty: 271.72, suggest sieving rational side
# size = 2.140e-018, alpha = 0.000, combined = 1.749e-014, rroots = 1
type: snfs
size: 267
skew: 1.0433
c5: 106
c0: 131
Y1: -290199866805246507499041077857400346603111731
Y0: 45493829629280918649510295477477883207043693313785856
rlim: 134000000
alim: 134000000
lpbr: 31
lpba: 31
mfbr: 91
mfba: 62
rlambda: 3.4
alambda: 2.7

16f test sieving yield, with Q in blocks of 2000:

60M   1.94
100M  1.96
150M  1.67
200M  1.67
250M  1.49

Suggesting a Q-range of 60-210M with a number of raw relations target = 260M.


The sextic:

n: 1030503235456803762016167471530641995037760274197968310988179590505950995510892101498055733828425469497568921009021217030573300085480516986964919661166160134908753817168257662257684691416593544651963249784543552924184296481653156578163
# 131^106+106^131, difficulty: 269.37, anorm: 2.70e+039, rnorm: -5.51e+050
# scaled difficulty: 271.25, suggest sieving rational side
# size = 6.893e-014, alpha = 0.669, combined = 1.663e-014, rroots = 0
type: snfs
size: 269
skew: 2.3346
c6: 106
c0: 17161
Y1: -360353741657835234074373091747178365988110336
Y0: 129087241933376588180389974363760340041
rlim: 225000000
alim: 225000000
lpbr: 32
lpba: 32
mfbr: 94
mfba: 64
rlambda: 3.5
alambda: 2.8

I have test sieving results of this deg 6 poly if anyone is interested. But it's a bust.

[axn]

Originally Posted by swellman

I test sieved both the deg 5 and deg 6 SNFS polys. Neither sieves all that efficiently, I assume due to the awkward coefficients. But the quintic clearly is more efficient in sieving - not even close. Maybe the sextic coefficients are just that much more awkward?

Did you try algebraic or rational side for sextic? If the latter, former might be worth a try.

EDIT:- Nevermind. The sextic is just ugly.

[swellman]

Originally Posted by axn

Did you try algebraic or rational side for sextic? If the latter, former might be worth a try.

I tried both sides. I usually look at 2/2, 3/2 and 2/3 LPs and test sieve each case on the -r and -a sides. Wasn’t a lot of difference but the rational side won out.

ETA: And yes, that is an ugly sextic!

[axn]

Originally Posted by swellman

I tried both sides. I usually look at 2/2, 3/2 and 2/3 LPs and test sieve each case on the -r and -a sides. Wasn’t a lot of difference but the rational side won out.

# 131^106+106^131, difficulty: 267.34
# 131^106+106^131, difficulty: 269.37

I don't get this part. IIUC, for getting the sextic, the number had to be multiplied by 131^2*106, whereas for the quintic, there is no such fiddling necessary. So the sextic difficulty should be 6 digits larger.

[swellman]

Originally Posted by axn

# 131^106+106^131, difficulty: 267.34
# 131^106+106^131, difficulty: 269.37

I don't get this part. IIUC, for getting the sextic, the number had to be multiplied by 131^2*106, whereas for the quintic, there is no such fiddling necessary. So the sextic difficulty should be 6 digits larger.

Output by Yafu. With SNFS it cranks out dozens of potential polynomials then downselects to the best three for test sieving. Perhaps B^2 can shed more light on the various steps of that Kabuki dance.

[VBCurtis]

On the quintic 16/31 as you wish to submit, I'd try bumping alim to 200M while leaving rlim alone. Both yield and sec/rel should improve slightly.
I'd also try mfba of 61 or 60- you might not gain speed, but it should reduce rels needed / final matrix size without a loss of sieve speed.

[swellman]

Originally Posted by VBCurtis

On the quintic 16/31 as you wish to submit, I'd try bumping alim to 200M while leaving rlim alone. Both yield and sec/rel should improve slightly.
I'd also try mfba of 61 or 60- you might not gain speed, but it should reduce rels needed / final matrix size without a loss of sieve speed.

I will try these. Should not adversely affect NFS@Home and it could improve efficiency all with no apparent downside. Thanks!