2022 - queue management for 15e_small - Page 16

unconnected · 2022-12-08, 22:41

QUEUED- see post 173 below

C173 from 785232:i11580 is ready for 15e_small queue:

Code:

n: 13412430154895484141862818606541773755250877526911498036451802532740000933585379389777742933827265930661389244830186926816851949173431811572533147379816321246734456628640519
# norm 9.395772e-17 alpha -6.459608 e 2.436e-13 rroots 5
skew: 11021533.76
c0: 29566114166770514028944559431151323845968
c1: 35325849076340889040277016898908978
c2: 8735032240722787163675795955
c3: -251146757830927681580
c4: -55178061598008
c5: 1876392
Y0: -1481961035514585249264277307725565
Y1: 2873252762515349443
rlim: 134000000
alim: 134000000
lpbr: 31
lpba: 31
mfbr: 62
mfba: 62
rlambda: 2.6
alambda: 2.6
type: gnfs
lss: 0

Suggesting sieving range is 40M-125M, I'll take the LA.

VBCurtis · 2022-12-09, 01:54

Quote:

Originally Posted by unconnected

C173 from 785232:i11580 is ready for 15e_small queue:
...
Suggesting sieving range is 40M-125M, I'll take the LA.

Check out the params in post #162, where we sieved a C172 in 36MQ versus your params that take 85MQ.

Q-range isn't a precise measure of speed (3LP takes longer per Q than 2LP, for one thing), but please consider testing my params against yours on your next job around 170-174 digits.

swellman · 2022-12-09, 03:04

Quote:

Originally Posted by VBCurtis

Check out the params in post #162, where we sieved a C172 in 36MQ versus your params that take 85MQ.

Q-range isn't a precise measure of speed (3LP takes longer per Q than 2LP, for one thing), but please consider testing my params against yours on your next job around 170-174 digits.

I’ll add that I performed the LA and it took only 49 hours on 24 cores, say ~1 week on a quad core. Fairly quick for a 32-bit job.

Though it did take until the 7th dependency before a factor popped up. Just bad luck I suppose.

VBCurtis · 2022-12-09, 07:45

Quote:

Originally Posted by swellman

I’ll add that I performed the LA and it took only 49 hours on 24 cores, say ~1 week on a quad core. Fairly quick for a 32-bit job.

Though it did take until the 7th dependency before a factor popped up. Just bad luck I suppose.

Adding 10-15M more relations would likely have reduced that matrix from 12.x million to 10-11M. I got a little too greedy trying to minimize Q-range, seems. If you filtered with a few million rels pending, that may have cost 0.5M dimensions on the matrix too; I wait till nearly all rels are in, I figure someone did the calcs I may as well use 'em.
I truly don't think the LP size of a job has much to do with matrix size- the difficulty of the job is what it is, whether we use 31 or 32 or 33 bit LPs, provided a small amount of "extra" sieving beyond the bare minimum that makes a matrix.

For instance, if you test-sieve a 31LP job, establish the Q-range, and then change only LP to 32 without changing any MFBs; then sieve the Q-range you planned to sieve at 31LP, I wager the resulting matrix would be *smaller* using 32LP. Try it some time! Submit an e-small job and just increase the LP by 1 and see what happens.

A 2LP job of the same size would likely make a matrix around 8-9M, so that's a fair argument for unconnected's params.

unconnected · 2022-12-10, 13:51

Thanks for your suggestions, guys. Unfortunately I've no time to test params till Monday.

swellman · 2022-12-10, 15:05

Quote:

Originally Posted by unconnected

Thanks for your suggestions, guys. Unfortunately I've no time to test params till Monday.

If you have no objection, I’ll test the suggested parameters tonight and report results here.

unconnected · 2022-12-10, 15:49

Of course, I've no objections. Thank you, Sean!

swellman · 2022-12-12, 03:41

Second set of parameters below queued as C173_785232_11580

Quote:

Originally Posted by swellman

If you have no objection, I’ll test the suggested parameters tonight and report results here.

I test sieved C173_785232_11580 on the -a side with the original parameters just to get some timings, and then again with the suggested parameters from post 162 of this thread.

Using the orginal posted job file:

Code:

n: 13412430154895484141862818606541773755250877526911498036451802532740000933585379389777742933827265930661389244830186926816851949173431811572533147379816321246734456628640519
# norm 9.395772e-17 alpha -6.459608 e 2.436e-13 rroots 5
skew: 11021533.76
type: gnfs
lss: 0
c0: 29566114166770514028944559431151323845968
c1: 35325849076340889040277016898908978
c2: 8735032240722787163675795955
c3: -251146757830927681580
c4: -55178061598008
c5: 1876392
Y0: -1481961035514585249264277307725565
Y1: 2873252762515349443
rlim: 134000000
alim: 134000000
lpbr: 31
lpba: 31
mfbr: 62
mfba: 62
rlambda: 2.6
alambda: 2.6

I got these results:

Code:

MQ       Norm_yield      speed (sec/rel)
40          2745              0.267
60          2688              0.250
80          2692              0.260
100         2660              0.237
125         2826              0.253

Which should generate ~230M raw relations.

Then I ran it again with the following parameters:

Code:

n: 13412430154895484141862818606541773755250877526911498036451802532740000933585379389777742933827265930661389244830186926816851949173431811572533147379816321246734456628640519
# norm 9.395772e-17 alpha -6.459608 e 2.436e-13 rroots 5
skew: 11021533.76
type: gnfs
lss: 0
c0: 29566114166770514028944559431151323845968
c1: 35325849076340889040277016898908978
c2: 8735032240722787163675795955
c3: -251146757830927681580
c4: -55178061598008
c5: 1876392
Y0: -1481961035514585249264277307725565
Y1: 2873252762515349443
rlim: 67000000
alim: 33500000
lpbr: 32
lpba: 32
mfbr: 62
mfba: 91
rlambda: 2.45
alambda: 3.75

Which resulted in the following:

Code:

MQ        Norm_yield      speed (sec/rel)
10           8237              0.097
15           8853              0.098
20           8873              0.105
25           8460              0.116
30           8284              0.119
40           7655              0.129
45           7496              0.104

Which, if we follow a similar strategy listed in post 162 (i.e. aim for 325M relations) and add 10-15M more relations in order to improve the matrix, we should use a sieving range for Q of 8-48M.

I say we go with the new parameters - they appear to be faster and proven to work.

unconnected · 2022-12-12, 12:47

Seems good for me. It will be great if we can get a matrix smaller than 10M.
For the reference, last 3 C173 jobs resulting in 8.0M, 7.2M and 9.4M matrices.

VBCurtis · 2022-12-12, 17:11

Quote:

Originally Posted by unconnected

Seems good for me. It will be great if we can get a matrix smaller than 10M.
For the reference, last 3 C173 jobs resulting in 8.0M, 7.2M and 9.4M matrices.

On the C172 Sean just did with these params, 340M rels got a 12M matrix. So, let's aim for 355M to get the matrix in the 10.x range; say an extra 5MQ?

swellman · 2022-12-12, 18:14

Quote:

Originally Posted by VBCurtis

On the C172 Sean just did with these params, 340M rels got a 12M matrix. So, let's aim for 355M to get the matrix in the 10.x range; say an extra 5MQ?

Ok I’ll add 5MQ to the sieving range.

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

2022-12-10, 13:51	#170
unconnected May 2009 Moscow, Russia 3×5×197 Posts	Thanks for your suggestions, guys. Unfortunately I've no time to test params till Monday.

2022-12-10, 15:49	#172
unconnected May 2009 Moscow, Russia 5613₈ Posts	Of course, I've no objections. Thank you, Sean!

2022-12-12, 12:47	#174
unconnected May 2009 Moscow, Russia 2955₁₀ Posts	Seems good for me. It will be great if we can get a matrix smaller than 10M. For the reference, last 3 C173 jobs resulting in 8.0M, 7.2M and 9.4M matrices.