View Single Post
Old 2020-09-13, 16:21   #124
charybdis
 
Apr 2020

11011012 Posts
Default

Sieving Q from 30M to 201M took 109.9M CPU-seconds, and produced:
Code:
Sun Sep 13 13:55:54 2020  commencing relation filtering
Sun Sep 13 13:55:54 2020  setting target matrix density to 110.0
Sun Sep 13 13:55:54 2020  estimated available RAM is 15845.8 MB
Sun Sep 13 13:55:54 2020  commencing duplicate removal, pass 1
Sun Sep 13 14:28:58 2020  found 92069979 hash collisions in 327933354 relations
Sun Sep 13 14:29:20 2020  commencing duplicate removal, pass 2
Sun Sep 13 14:35:46 2020  found 109446093 duplicates and 218487261 unique relations
Sun Sep 13 14:35:46 2020  memory use: 2387.0 MB
Sun Sep 13 14:35:46 2020  reading ideals above 200998912
Sun Sep 13 14:35:46 2020  commencing singleton removal, initial pass
Sun Sep 13 14:52:41 2020  memory use: 5512.0 MB
Sun Sep 13 14:52:42 2020  reading all ideals from disk
Sun Sep 13 14:53:09 2020  memory use: 3752.4 MB
Sun Sep 13 14:53:14 2020  commencing in-memory singleton removal
Sun Sep 13 14:53:20 2020  begin with 218487261 relations and 200826425 unique ideals
...
Sun Sep 13 16:09:52 2020  matrix is 18797898 x 18798123 (8118.4 MB) with weight 2166084557 (115.23/col)
Sun Sep 13 16:09:52 2020  sparse part has weight 1940221126 (103.21/col)
Sun Sep 13 16:09:52 2020  using block size 8192 and superblock size 884736 for processor cache size 9216 kB
Sun Sep 13 16:10:46 2020  commencing Lanczos iteration (6 threads)
Sun Sep 13 16:10:46 2020  memory use: 7721.7 MB
Sun Sep 13 16:11:38 2020  linear algebra at 0.0%, ETA 171h51m
Higher lims don't seem to require more unique relations to build a matrix, though they do increase the duplication rate because more of the sieving is below lim1. The matrix hasn't got much larger either.
Taking into account the poly scores, speedup relative to the lower lims is about 5%.

I'm going to do the c183 from 4-3_443 next, as Sean's test-sieving showed that GNFS ought to be slightly faster than SNFS. I'll use lpb 32/32, mfb 60/90.
charybdis is offline   Reply With Quote