mersenneforum.org Reserved for MF - Sequence 3366
 Register FAQ Search Today's Posts Mark Forums Read

 2021-08-29, 02:48 #496 frmky     Jul 2003 So Cal 22·547 Posts Overnight last night I ran stage 1 on 2560 curves with B1=85e7 using the new gpu ecm using CGBN. I have stage 2 on those distributed across 8 nodes right now, and I'm running another 2560 stage 1's on the gpus.
 2021-09-01, 17:17 #497 charybdis     Apr 2020 49910 Posts Is there still a need for more ECM here, or is Greg doing all the 850M curves?
 2021-09-01, 19:33 #498 kruoli     "Oliver" Sep 2017 Porta Westfalica, DE 65410 Posts Besides the curves from frmky in #496 I do not know of anyone other than me who did some. We overshot the goal suggested by VBCurtis with 30k curves @ 260M, so I guess we need less 850M curves than he suggested. If frmky's curves are done, together with mine we are only at ~7k curves here. So there could certainly be done more, if you like. Last fiddled with by kruoli on 2021-09-01 at 19:35 Reason: Clarifications.
 2021-09-01, 20:01 #499 swellman     Jun 2012 3,203 Posts I have completed 6432 of my committed 10000 curves @B1=260M. Seeing we’ve already overshot the target amount of ECM at t60, would it be better for me to cease that work and do say 1000 curves @B1=850M? Sean
 2021-09-01, 20:38 #500 frmky     Jul 2003 So Cal 22×547 Posts I have run a total of 10240 curves at B1=850M with no factor found. Last fiddled with by frmky on 2021-09-01 at 20:39
2021-09-01, 21:09   #501
swellman

Jun 2012

1100100000112 Posts

Quote:
 Originally Posted by frmky I have run a total of 10240 curves at B1=850M with no factor found.
And I’m done with ECM! [Drops mic]

ETA: I can help with the poly search if needed.

Last fiddled with by swellman on 2021-09-01 at 22:25

 2021-09-02, 08:55 #502 kruoli     "Oliver" Sep 2017 Porta Westfalica, DE 65410 Posts Since we were discussing whether to use CADO or NFS@home, does it make a difference which poly selection to use? (IIRC, CADO yields better polynomials on CPU but has no GPU port.) With a CADO server, it would automagically distribute polyselect working units, correct? For a few c140-c150, I experimented in the past with polynomial selecting on msieve and using that result in CADO, resulting every time in an LA phase taking around 50 % of the computation time, so decided to stop experimenting until I get a better understanding of everything. We are looking at around 1000 core days of polyselect, right? For about five days already, I have been running GPU polyselect on a lousy 1030. The resulting polynomial was abominable. I guess the badness of the poly above at least does make sense... Code: R0: -1914165953547080696287480071699336060476 R1: 4996785037918941149 A0: 5557784762717745111741234027368983334161163620943009595 A1: 709924490299956872317450962824671796034061769 A2: -176580573731696825434979940746709615 A3: -839292886902290685831381 A4: 483459688232660 A5: 5100 skew 20143410012.87, size 8.676e-20, alpha -9.773, combined = 3.745e-15 rroots = 5
 2021-09-02, 11:06 #503 swellman     Jun 2012 3,203 Posts I don’t think the siever “cares” where the poly was generated. For a c201 I would use CADO for a degree 5 poly search. Sieve it by ggnfs or CADO, either will work. If you’re looking for assistance with the search, I hereby reserve a range of 5-10M using CADO.
 2021-09-02, 12:58 #504 bur     Aug 2020 79*6581e-4;3*2539e-3 3·7·19 Posts I could also do a 1M range. Should the same parameters be used as for the c220 search just with an nq of 5^n?
2021-09-02, 13:28   #505
charybdis

Apr 2020

499 Posts

I've done 2400 curves at 850M, no factor.

I think the CADO default of P=10M for c200 is probably too big: Curtis's optimized c195 polyselect parameters have P=3M.
I'll do 0-5M with P=5M, nq=15625, incr=420, sopteffort=10, ropteffort=100.

Quote:
 Originally Posted by kruoli For a few c140-c150, I experimented in the past with polynomial selecting on msieve and using that result in CADO, resulting every time in an LA phase taking around 50 % of the computation time, so decided to stop experimenting until I get a better understanding of everything.
Off-topic, but I'm guessing this was either on a big machine or sieving across multiple machines? Sieving is inherently parallel whereas LA isn't, so LA speed doesn't scale with the number of threads in the way that sieving does. Therefore the (LA time)/(sieving time) ratio increases as you add threads. The best solution is to make CADO oversieve a bit to reduce the matrix size; you can do this by adding tasks.filter.required_excess = 0.05 to the params file (experiment with different values to see what's best for your system).

2021-09-02, 14:18   #506
swellman

Jun 2012

3,203 Posts

Quote:
 Originally Posted by bur I could also do a 1M range. Should the same parameters be used as for the c220 search just with an nq of 5^n?
I’m using nq=5^6, seems to give good results.

As charybdis points out, P should probably be lower. I’m using P=8M, but that’s maybe a touch high.

Last fiddled with by swellman on 2021-09-02 at 14:18

 Similar Threads Thread Thread Starter Forum Replies Last Post sweety439 And now for something completely different 17 2017-06-13 03:49 RichD Aliquot Sequences 36 2013-11-29 07:03 petrw1 Lone Mersenne Hunters 82 2010-01-11 01:57 roger Puzzles 16 2006-10-18 19:52 Citrix Puzzles 5 2005-09-14 23:33

All times are UTC. The time now is 08:31.

Thu Oct 21 08:31:44 UTC 2021 up 90 days, 3 hrs, 1 user, load averages: 1.68, 1.29, 1.24