mersenneforum.org Reserved for MF - Sequence 4788
 Register FAQ Search Today's Posts Mark Forums Read

 2021-08-31, 16:51 #3092 bur     Aug 2020 79*6581e-4;3*2539e-3 1100011112 Posts If it makes sense I could also test-sieve, though only with 10 cores. If so, which would be a good set of parameters?
 2021-08-31, 18:52 #3093 charybdis     Apr 2020 22·53 Posts There's nothing stopping you from test-sieving, but it wouldn't be useful for me because I'm comparing speeds as well as yields for lots of different parameter choices, and for speeds to be consistent I have to run everything on the same machine.
 2021-08-31, 20:56 #3094 bur     Aug 2020 79*6581e-4;3*2539e-3 3×7×19 Posts Ok, then I won't get into it.
 2021-08-31, 23:33 #3095 charybdis     Apr 2020 7648 Posts Test-sieving done, at least for today. I only used I=16, lpb 34/35, and given the yields I don't expect anything smaller than this to be better. Even if 35/35 does sieve faster, it will make postprocessing more of a pain. The machine I used only has 16GB RAM so I can't test A=32 which requires ~26GB, but I don't think we're at a size where that would be clearly better. Raising ncurves0 from 25 doesn't find any more relations. It could possibly be dropped a bit, but 2LP cofactorization is a tiny contribution to the total time so the effect would be negligible. ncurves1=21 does miss a few relations, but it's just a fraction of a percent, so with the extra 3LP cofactorization time I don't think there's any need to change this from the default. There is a clear benefit of several percent from using -adjust-strategy 2, as seems to be standard with large jobs. The price paid for this is that memory use with I=16 is 13GB rather than 10GB, but I assume this won't be a problem. Higher lims = more yield and slower sieving, and with the added complication of the duplication rate depending on the lim choice, it's not easy to measure the effects with test-sieving. 500M/800M probably isn't optimal but it'll be close enough. mfb0 choice is always complicated by the fact that higher mfb0 leads to a small increase in the number of relations required, and we don't have a good handle on this. This makes 68 less attractive, but in the interests of keeping yield high I stuck with 67 for the main test. mfb1 can go quite high and still show a decent improvement in sec/rel. There doesn't seem to be any problem with going as high as 102. I test-sieved the 500M/800M, 67/102 combination at various intervals between 200M and 2500M. Sec/rel worsened from ~0.7 to ~1.5 over this range. If the aim is 3G raw relations, the range 250M-2000M looks about right, and I estimate this will take only(!) ~105 thread-years on the machine I used. This figure seems suspiciously low even given that the threads in question are fast, but I can't find a mistake anywhere, so I have to believe it. Hopefully this doesn't mean Ryan's in for a nasty surprise with the duplication rate! Last fiddled with by charybdis on 2021-08-31 at 23:37
 2021-09-01, 02:45 #3096 VBCurtis     "Curtis" Feb 2005 Riverside, CA 19×263 Posts I appreciate the detail in the writeup- for e.g. Bur who sees a future in test-sieving, that post is a nice place to start. Here's the conclusion for Ryan: Code: tasks.lim0 = 500000000 tasks.lim1 = 800000000 tasks.lpb0 = 34 tasks.lpb1 = 35 tasks.sieve.mfb0 = 67 tasks.sieve.mfb1 = 102 tasks.sieve.ncurves0 = 25 tasks.sieve.ncurves1 = 21 tasks.I = 16 tasks.qmin = 250000000 tasks.sieve.qrange = 5000 tasks.sieve.adjust_strategy = 2 tasks.sieve.rels_wanted = 3000000000 If memory is plentiful, consider running A=32 for the first ~500M relations for the extra yield at small Q, and then changing down to I=16. That should help with the duplicate-relations rate, too- making it more likely to get 2G unique relations before Q=2000M.
 2021-09-01, 03:02 #3097 charybdis     Apr 2020 22·53 Posts I'd make one small change for practical purposes, which is to set rels_wanted artificially high (say 4G rather than 3G) as the intention is presumably to run msieve filtering. We don't want CADO filtering to be triggered, for a couple of reasons: 1. There is a bug whereby sieving with adjust_strategy = 2 can produce a relation with a (square) composite factor. This is extremely rare, but in 3G relations there's a decent chance it will happen, and unlike msieve, I don't think CADO filtering handles it well. 2. Even if this doesn't happen, CADO filtering on a job this big likely requires hundreds of GB of memory. Setting a large value of rels_wanted has the added benefit that ETA values are more accurate early in the sieving while the yield is high
2021-09-01, 17:19   #3098
bur

Aug 2020
79*6581e-4;3*2539e-3

1100011112 Posts

Quote:
 for e.g. Bur who sees a future in test-sieving, that post is a nice place to start.
That sound like test-sieving is a career-option. Professional test-siever.

But seriously, thanks, I always enjoy these posts.

2021-09-12, 18:56   #3099
ryanp

Jun 2012
Boulder, CO

52×13 Posts

Quote:
 Originally Posted by VBCurtis I appreciate the detail in the writeup- for e.g. Bur who sees a future in test-sieving, that post is a nice place to start. Here's the conclusion for Ryan:
I'm sieving now. Thanks to those who found the poly and params!

Will give an update in a few days.

 2021-09-14, 13:58 #3100 ryanp     Jun 2012 Boulder, CO 52×13 Posts Update: I'm at 555M uniques now. I currently seem to be pulling about 250M unique (not total) relations per day, so hopefully another 6-10 days of sieving should be enough, if the rate keeps up.
 2021-09-14, 16:56 #3101 bur     Aug 2020 79*6581e-4;3*2539e-3 3·7·19 Posts How can the percentage of uniques be found during sieving? By doing a manual filtering step on the relations? Impressive progress anyway, about the same timeframe I need for a C167...
2021-09-14, 20:21   #3102
EdH

"Ed Hall"
Dec 2009

32×5×89 Posts

Quote:
 Originally Posted by bur How can the percentage of uniques be found during sieving? By doing a manual filtering step on the relations? Impressive progress anyway, about the same timeframe I need for a C167...
remdups4 is what I use. It can be found in the ggnfs package. (ggnfs/contrib/remdups/remdups4.c)

Edit: Depending on which sieving package you are using, you will need to gather all the relations into a single file and:
Code:
[z]cat <file> | ./remdups4 1000 ><uniquerels>

Last fiddled with by EdH on 2021-09-14 at 20:24

 Similar Threads Thread Thread Starter Forum Replies Last Post kar_bon Aliquot Sequences 136 2021-10-21 16:17 RichD Aliquot Sequences 476 2021-10-04 20:47 RichD Aliquot Sequences 524 2021-09-06 21:00 prism019 GPU to 72 6 2020-09-21 22:11 petrw1 Lone Mersenne Hunters 82 2010-01-11 01:57

All times are UTC. The time now is 12:02.

Fri Oct 22 12:02:24 UTC 2021 up 91 days, 6:31, 1 user, load averages: 1.36, 1.33, 1.29