20210720, 13:19  #12  
Apr 2020
17·29 Posts 
Quote:
I'd have used roughly c135 parameters for both jobs given the E scores. Quote:
If the rels_wanted figure was off for the c135 parameters, maybe it was off for the c130 parameters too. That still doesn't explain the discrepancy between the two numbers. By the way, it is essential to add tasks.sieve.adjust_strategy = 2 whenever you use tasks.A = [even number] rather than tasks.I. There should be a noticeable speedup. 

20210720, 16:20  #13  
Aug 2020
79*6581e4;3*2539e3
616_{8} Posts 
Quote:
Other than that, I don't remember the exact point of this increase but it was quite sudden. I found a possible explanation though. I don't do every number by SNFS since most are factored by easy ECM. So it might have been that I compared a number with an exponent divisible by 5 to one that was == 4 (mod 5). SNFS179 number 1281979*2^594+1 with polynomial 1281979*2^4*(2^118)^5+1 has an Escore of 3.365E11 while the SNFS175 number 1281979*2^580+1 yields an Escore of 5.747E11. I am not sure how strongly that affects sieving times, but the former Escore is 42% lower, so it might be quite significant? Add to that the problem with the missing adjust_strategy and it might explain everything. Last fiddled with by bur on 20210720 at 16:21 

20210720, 17:42  #14  
Apr 2020
17×29 Posts 
Quote:


20210720, 18:10  #15 
Aug 2020
79*6581e4;3*2539e3
2×199 Posts 
Alright, thanks. The tasks.sieve.adjust_strategy = 2, is there a general rule if it's useful for SNFS or not? And if I was to test and change this back to 0, I should probably also revert to tasks.I? From the manual either 13 or 14 would correspond to A = 26, correct?

20210720, 18:59  #16 
Apr 2020
17×29 Posts 
tasks.I = n corresponds to tasks.A = 2n1, so A=26 is between I=13 and I=14.
If you're using an even value of A, you should use adjust_strategy = 2 whatever the circumstances. For tasks.I (in other words, odd values of A), VBCurtis's testing showed that adjust_strategy = 2 is advantageous for GNFS from around 125 digits upwards, though the difference in speed is very small. I don't think much SNFS testing has been done but I'd expect the dynamics to be similar, i.e. use strategy 2 for SNFS jobs of difficulty equivalent to GNFS125 or higher. 
20210721, 05:45  #17 
Aug 2020
79*6581e4;3*2539e3
2·199 Posts 
Yes, according to CADO manual A can be used to finetune that parameter. So for integer values of I (or odd values of A) it doesn't matter which one is used?
This 188 digit 2^602*1281979+1 took 9:20 for sieving, even with adjust_strategy. So I guess it is just a peculiarity of these SNFS numbers that the Escores are low and sieving takes long. Btw, is the only difference between GNFS and SNFS that the latter has a simple polynomial? Earlier I was under the impression that SNFS was having a different sieving process, but it doesn't seem so. Or does cado recognize the SNFS type polynomial and internally switches to "SNFS"? 
20210721, 11:37  #18 
Apr 2020
17×29 Posts 
The advantage for adjust_strategy=2 with integer I seems to get greater for larger jobs. I don't know why this happens, but it does. So if you're doing a big job then you definitely will want strategy 2: saving 5% of sieving time on a c180 GNFS is absolutely worth it.
I don't know how your timings for GNFS and other SNFS jobs compare, so I can't tell what counts as slow. If you're meaning that these SNFS jobs take longer than the "rules of thumb" for SNFSGNFS conversion indicate, then yes  that's because of the large c5 coefficient. There is no difference between the SNFS and GNFS algorithms once we've got a polynomial. SNFS just means the special form of the number allows a polynomial to be easily found, usually with small coefficients. (Historical note: in the early days, there was more of a difference, because SNFS was initially only developed for numbers of the form b^n + c where the algebraic number theory is easier.) 
20210722, 06:38  #19 
Aug 2020
79*6581e4;3*2539e3
2×199 Posts 
Ok thanks, good to know.

Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
I found the primality test, there seems to be no composite numbers that pass the test  sweety439  sweety439  7  20200211 19:49 
Modifying the Lucas Lehmer Primality Test into a fast test of nothing  Trilo  Miscellaneous Math  25  20180311 23:20 
Advantage of lattice sieve over line sieve  binu  Factoring  3  20130413 16:32 
Double check LL test faster than first run test  lidocorc  Software  3  20081203 15:12 
A primality test for Fermat numbers faster than Pépin's test ?  T.Rex  Math  0  20041026 21:37 