2022 - queue management for 15e_small - Page 7

swellman · 2022-05-24, 17:26

Quote:

Originally Posted by chris2be8

I've ECMed (51^169-1)/(51^13-1) to nearly t60, so it's ready for SNFS. But test sieving suggests it's too big for 15e_small, I could only get a nearly high enough yield with lpb[ar]: 32 and 3 large primes. Which would probably get a very tough matrix.

If I decide to send it to 15e for larger [ar]lim and sieving range how would that affect the matrix size? And are there any recommended lim's (IIRC the siever slows with lims just over 134M)?

This job almost certainly should run on 15e. Remember the lims on 15e are limited by Greg to 200M each, or rather rlim+alim =< 400M. You can use the trick where the lim on the sieving side = 134M, the other lim value = 266M. Improves yield a touch. Not sure how or if it affects matrix size. 3 LPs will likely help much more of course, again without comment on the resulting matrix.

Another strategy you could try is running this job on 15e_small with all its constraints and then sieve locally until you can get the job into LA.

VBCurtis · 2022-05-24, 18:04

3 large prime jobs produce matrices 10-20% larger than jobs with 2LP on both sides, when comparing the same-size job (both SNFS and GNFS). I usually go for 5-10% more raw relations when stepping from 2LP on both sides to 3LP on one side, to try to mitigate this. For me, when test-sieving I want to see 3LP at least 10% faster than 2LP to make up for these effects.

Larger lims do make a larger matrix, but I haven't noticed any obvious scaling to advise, such as 200/200 lim's making a 5% larger matrix than 134/134. I can say the effect must not be too big, else the effect would stick out more clearly in our data- maybe 5% or less.

If 3LP is on the sieve side, keeping sieve-side lim at 134M and pushing the other side to 266M maximizes the yield possible under Greg's "lims add to 400M" guidance, a rule that caps memory use of each siever process. Another trick that adds yield without (notably) larger matrix is to add an LP to the 2LP side, 32/33 instead of 32/32. I've run a bunch of jobs with 33/64 or 33/65 on the non-sieve-side, and required number of relations to get the same-size matrix as a 32/32 job only rises by 30-35%; I usually find yield goes up by more than 35%, so it seems a clear win.

EDIT: I didn't quite answer all your questions. Sieve range is an effect rather than a cause- the difficulty of a job causes sieve range and matrix size both. I don't think smaller sieve ranges lead to smaller matrices for a given job, though smaller Q-max to Q-min ratio seems to produce fewer duplicate relations. I test-sieve to minimize total sieve time, though that's usually also when Q-range is minimized.
As for recommended lim's, if I'm going above 134/134 I jump to 134/200 or 170/170 at minimum- any closer to 134 and you may as well stay with 134/134 because of the speed hit you mentioned. I test 134/200 against 134/266 generally, but something like 180/220 or 200/200 might be best if 2LP on both sides- the unbalanced lims work better on 3LP jobs with the small lim on the 3LP side, since the siever will pick up most of the relations that would have been lim 200M & 2LP with 134M and 3LP, at least when one prime is within [134M, 200M].

chris2be8 · 2022-05-25, 16:02

Thanks for the advice. I'll test sieve lims of 200M/200M against 133M/266M and see which gets the best yield. With luck I'll be able to post a job tomorrow.

I wish I knew what all the output from the siever means. It looks as if it should mean something, but most of it doesn't mean anything to me.

unconnected · 2022-05-25, 23:07

QUEUED AS C184_933436_12610

C184 from 933436:i12610 is ready for GNFS:

Code:

n: 4876630858212349810691213611925050755952048334305408903910300951464902067092316845896254651320465422244744087225834090890961772198806089897040721540280769008612228589646966410548502467
# skew 50681508.57, size 5.067e-018, alpha -8.241, combined = 4.711e-014 rroots = 5
skew: 50681508.57
c0: -110483875328199321559431815261460398065873320
c1: 64383384366687242237371786946650854654
c2: -761645235654446236920932275627
c3: -82055981426510868277147
c4: -223481580525450
c5: 8164800
Y0: -324241054145458129729151114244781903
Y1: 393459119910895718201687
rlim: 266000000
alim: 134000000
lpbr: 32
lpba: 32
mfbr: 64
mfba: 94
rlambda: 2.7
alambda: 3.6
lss: 0
type: gnfs

Thanks swellman for the poly!
Suggesting sieving range is 30M-180M.

swellman · 2022-05-26, 12:11

Quote:

Originally Posted by unconnected

C184 from 933436:i12610 is ready for GNFS:

Code:

n: 4876630858212349810691213611925050755952048334305408903910300951464902067092316845896254651320465422244744087225834090890961772198806089897040721540280769008612228589646966410548502467
# skew 50681508.57, size 5.067e-018, alpha -8.241, combined = 4.711e-014 rroots = 5
skew: 50681508.57
c0: -110483875328199321559431815261460398065873320
c1: 64383384366687242237371786946650854654
c2: -761645235654446236920932275627
c3: -82055981426510868277147
c4: -223481580525450
c5: 8164800
Y0: -324241054145458129729151114244781903
Y1: 393459119910895718201687
rlim: 266000000
alim: 134000000
lpbr: 32
lpba: 32
mfbr: 64
mfba: 94
rlambda: 2.7
alambda: 3.6
lss: 0
type: gnfs

Thanks swellman for the poly!
Suggesting sieving range is 30M-180M.

One issue - jobs on 15e_small are limited to both lims = 134M and the sieving range Qmax-Q(0) =< 170M. You’ve met the second criteria but not the first. (3 LPs are fine.)

I have dropped rlim to 134M. Suggest you test sieve again to generate a new yield curve. Hoping 30-200M will be your new sieving range.

ETA: if this job absolutely won’t fit within 15e_small then we can just move it over to 15e.

unconnected · 2022-05-26, 14:19

Quote:

Originally Posted by swellman

One issue - jobs on 15e_small are limited to both lims = 134M and the sieving range Qmax-Q(0) =< 170M. You’ve met the second criteria but not the first. (3 LPs are fine.)

I have dropped rlim to 134M. Suggest you test sieve again to generate a new yield curve. Hoping 30-200M will be your new sieving range.

ETA: if this job absolutely won’t fit within 15e_small then we can just move it over to 15e.

According to my test-sieve increased rlim gives about 10% additional relations, so setting alim=rlim=134M and sieve range to 170M (30M-200M) should be OK.

swellman · 2022-05-26, 14:51

Quote:

Originally Posted by unconnected

According to my test-sieve increased rlim gives about 10% additional relations, so setting alim=rlim=134M and sieve range to 170M (30M-200M) should be OK.

Done.

swellman · 2022-06-09, 15:29

QUEUED AS 7p4_338

7+4,338 is a HCN just finishing ECM. It is a SNFS 263 job to be run on 15e_small.

Code:

n: 40163610219963325261908698168315725543710527205495906950565894692963048457252960264218821247847727257524935332555526378042942452539538527605436322858575428931783709877273726423041462668170033957492670534887776342052666307717
skew: 1.0000
type: snfs
size: 263
c6: 1
c5: -1
c4: -5
c3: 4
c2: 6
c1: -3
c0: -1
Y1: -42277452950578284263485622772148731904
Y0: 88124787089743477593997388358336765364597217
rlim: 134000000
alim: 134000000
lpbr: 31
lpba: 31
mfbr: 91
mfba: 62
rlambda: 3.4
alambda: 2.7

Results of test sieving on the rational side with Q in blocks of 1000:

Code:

MQ        Norm_yield      Speed (sec/rel)
35           2093              0.220
50           1831              0.290
75           1741              0.297
100          1551              0.292
150          1343              0.404
205          1102              0.453

Suggesting a sieving range for Q of 35-205M to generate 255M raw relations.

I will take the LA.

swellman · 2022-06-14, 14:25

QUEUED AS 8p3_292

8+3,292 is a HCN now ready for SNFS on 15e_small.

Code:

n: 267136507979524605470044357395964682249086243652588091243912977344455967651988338045265512944693533079436212360808029902035582310353325504558218309464460120217516364304443337051340619644876368298145800332116647634660653788114592729385267437199593
skew: 0.3467
type: snfs
size: 264
c6: 576
c0: 1
Y1: -239299329230617529590083
Y0: 44601490397061246283071436545296723011960832
rlim: 134000000
alim: 134000000
lpbr: 31
lpba: 31
mfbr: 91
mfba: 62
rlambda: 3.4
alambda: 2.7

Results of test sieving on the rational side with Q in blocks of 1000:

Code:

MQ       Norm_yield      Speed (sec/rel)
35          1875              0.300
50          1691              0.339
75          1531              0.361
100         1389              0.403
140         1257              0.595
180         1203              0.507
205         1060              0.629

Suggesting a sieving range for Q of 35-205M to generate 235M raw relations. I will take the LA.

swellman · 2022-06-17, 10:18

QUEUED AS 7p5_317

7+5,317 is a HCN ready for SNFS on 15e_small.

Code:

n: 8005250922964400623653409089801614163844341926791193708170502887552253315908849909972739725496845462858419218045491332116038813116004297778170450465664390084878683606503403611971956761968587663149
skew: 1.0577
type: snfs
size: 267
c6: 5
c0: 7
Y1: -11102230246251565404236316680908203125
Y0: 616873509628062366290756156815389726793178407
rlim: 134000000
alim: 134000000
lpbr: 32
lpba: 32
mfbr: 94
mfba: 64
rlambda: 3.5
alambda: 2.8

Results of test sieving on the rational side with Q in blocks of 1000:

Code:

MQ        Norm_yield      Speed (sec/rel)
35           3105             0.220
50           3184             0.204
75           2724             0.228
100          2682             0.256
140          2237             0.343
180          2025             0.333
205          2007             0.334

Suggesting a sieving range for Q of 35-205M to generate 425M relations. I will do the LA.

swellman · 2022-06-21, 12:08

QUEUED AS 11p2_257

11+2,257 from the HCN project is ready for SNFS on 15e_small.

Code:

n: 3558673052264700164798309434496977032392885183395428816268310259594883831061052005721913652349571099501066977072009019675667943856066053787793760072797782324893647284132923852503247722872061024557
type: snfs
size: 267
skew: 1.3286
c6: 2
c0: 11
Y1: -8796093022208
Y0: 602400691612421918536387328824478011400331731
rlim: 134000000
alim: 134000000
lpbr: 32
lpba: 32
mfbr: 94
mfba: 64
rlambda: 3.5
alambda: 2.9

Results of test sieving on the rational side with Q in blocks of 1000:

Code:

MQ       Norm_yield      Speed (sec/rel)
35          3194              0.227
50          2906              0.275
75          2756              0.296
100         2484              0.342
140         2227              0.382
180         2015              0.415
205         1935              0.443

Suggesting a sieving range for Q of 35-205M to generate 410M raw relations. I will do the LA.

2022-05-24, 18:04	#68
VBCurtis "Curtis" Feb 2005 Riverside, CA 2×3²×17×19 Posts	3 large prime jobs produce matrices 10-20% larger than jobs with 2LP on both sides, when comparing the same-size job (both SNFS and GNFS). I usually go for 5-10% more raw relations when stepping from 2LP on both sides to 3LP on one side, to try to mitigate this. For me, when test-sieving I want to see 3LP at least 10% faster than 2LP to make up for these effects. Larger lims do make a larger matrix, but I haven't noticed any obvious scaling to advise, such as 200/200 lim's making a 5% larger matrix than 134/134. I can say the effect must not be too big, else the effect would stick out more clearly in our data- maybe 5% or less. If 3LP is on the sieve side, keeping sieve-side lim at 134M and pushing the other side to 266M maximizes the yield possible under Greg's "lims add to 400M" guidance, a rule that caps memory use of each siever process. Another trick that adds yield without (notably) larger matrix is to add an LP to the 2LP side, 32/33 instead of 32/32. I've run a bunch of jobs with 33/64 or 33/65 on the non-sieve-side, and required number of relations to get the same-size matrix as a 32/32 job only rises by 30-35%; I usually find yield goes up by more than 35%, so it seems a clear win. EDIT: I didn't quite answer all your questions. Sieve range is an effect rather than a cause- the difficulty of a job causes sieve range and matrix size both. I don't think smaller sieve ranges lead to smaller matrices for a given job, though smaller Q-max to Q-min ratio seems to produce fewer duplicate relations. I test-sieve to minimize total sieve time, though that's usually also when Q-range is minimized. As for recommended lim's, if I'm going above 134/134 I jump to 134/200 or 170/170 at minimum- any closer to 134 and you may as well stay with 134/134 because of the speed hit you mentioned. I test 134/200 against 134/266 generally, but something like 180/220 or 200/200 might be best if 2LP on both sides- the unbalanced lims work better on 3LP jobs with the small lim on the 3LP side, since the siever will pick up most of the relations that would have been lim 200M & 2LP with 134M and 3LP, at least when one prime is within [134M, 200M]. Last fiddled with by VBCurtis on 2022-05-24 at 18:13

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Queue management for 14e queue	VBCurtis	NFS@Home	170	2023-01-02 15:27
2022 Queue management of 15e	swellman	NFS@Home	186	2022-12-27 12:53
Queue management for 16e queue	VBCurtis	NFS@Home	154	2022-12-23 21:35
Queue management for e_small and 15e queues	VBCurtis	NFS@Home	254	2022-01-02 01:59
Improving the queue management.	debrouxl	NFS@Home	10	2018-05-06 21:05

2022-05-25, 16:02	#69
chris2be8 Sep 2009 5·491 Posts	Thanks for the advice. I'll test sieve lims of 200M/200M against 133M/266M and see which gets the best yield. With luck I'll be able to post a job tomorrow. I wish I knew what all the output from the siever means. It looks as if it should mean something, but most of it doesn't mean anything to me.