110.4M CPU-seconds of sieving, from 30M to 207M, gave:
Code:
Fri Sep 25 13:16:07 2020 commencing relation filtering
Fri Sep 25 13:16:07 2020 setting target matrix density to 110.0
...
Fri Sep 25 13:59:11 2020 found 108139037 hash collisions in 408503325 relations
Fri Sep 25 13:59:32 2020 commencing duplicate removal, pass 2
Fri Sep 25 14:08:03 2020 found 121841205 duplicates and 286662120 unique relations
Fri Sep 25 14:08:03 2020 memory use: 2387.0 MB
Fri Sep 25 14:08:04 2020 reading ideals above 207028224
Fri Sep 25 14:08:04 2020 commencing singleton removal, initial pass
Fri Sep 25 14:30:25 2020 memory use: 6024.0 MB
Fri Sep 25 14:30:26 2020 reading all ideals from disk
Fri Sep 25 14:31:11 2020 memory use: 5114.7 MB
Fri Sep 25 14:31:19 2020 commencing in-memory singleton removal
Fri Sep 25 14:31:27 2020 begin with 286662120 relations and 262594881 unique ideals
...
Fri Sep 25 16:00:53 2020 matrix is 19772369 x 19772594 (8473.1 MB) with weight 2251213167 (113.86/col)
Fri Sep 25 16:00:53 2020 sparse part has weight 2023439071 (102.34/col)
Fri Sep 25 16:00:53 2020 using block size 8192 and superblock size 884736 for processor cache size 9216 kB
Fri Sep 25 16:01:50 2020 commencing Lanczos iteration (6 threads)
Fri Sep 25 16:01:50 2020 memory use: 8099.3 MB
Fri Sep 25 16:02:44 2020 linear algebra at 0.0%, ETA 188h45m
Quote:
Originally Posted by VBCurtis
There's a strong chance it's just a lucky polynomial with better-than-avg duplication rate. I could see a small improvement in unique-to-raw ratio if this run needed a smaller Q range (i.e. much better yield), but otherwise I wager it's just luck.
|
The polynomial score actually isn't great: 4.768e-14 for a large c183, compared to 5.544e-14 for the c184 I ran previously. It did get an unusually large boost of around 3% from the re-scoring that CADO runs at the end of polyselect, which took it up from 4th to 1st in the list of best polys - though even adding 3% to 4.768e-14 only gets you up to 4.911e-14.
If we assume that this poly is indeed ~13% worse than the c184 poly, then 32/32 is a clear win. But if the low duplication rate wasn't picked up by the poly score, then this might not be a fair comparison, so I think another run with identical parameters is in order.