20050916, 13:49  #12 
May 2005
2^{2}·11·37 Posts 
So testing k*2^nb for k=736320585 is not the optimal way to use LLR?
Last fiddled with by Cruelty on 20050916 at 13:49 
20050916, 15:27  #13  
May 2004
FRANCE
267_{16} Posts 
Using LLR with big k's
Quote:
If b == 1, LLR will do a proving test ; k being large, gwnum will work in generic mode in both cases, so you will not get the IBDWT performances, but I don't know if a faster program is presently available (try Openpfgw, but it also uses the gwnum code...). Jean 

20050918, 19:11  #14 
Jun 2003
3105_{8} Posts 
Jean,
which is the fastest k to test, in terms of speed. Are all k under 2^20 the same speed? What is the difference in speed between a k under 2^20 and a k between 2^20 and 2^21? Citrix 
20050919, 07:40  #15  
May 2004
FRANCE
3·5·41 Posts 
Quote:
1) The speed is determined by the FFT length necessary to process a number N of given bit length. 2) According to the k size, the gwnum code uses three different algorithms to do multiplications : pure IBDWT, Zero paded IBDWT, generic mode. 3) The pure IBDWT algorithm can only be used with k values from 1 to around 2^20 ; it is the most efficient, because it requires the smallest FFT length, and makes the modular reduction totally free. Nevertheless, the FFT length for a given size of N increases smoothly by a factor of 2 when k goes from 1 to 2^20, then, the Zero padded algorithm becomes the better. 4) The Zero padded IBDWT can be used for k values up to around 2^48, and the performances continue to decrease smoothly. 5) For higher k values, the generic mode is used, and the speed for given size of N is around three times smaller than the IBDWT one. I hope this rather involved explanation will satisfy you. Jean 

20050919, 07:43  #16 
Jun 2003
3×5×107 Posts 
The smaller the k the faster it will be? Did I get it right?
Citrix 
20050919, 08:25  #17  
May 2004
FRANCE
3·5·41 Posts 
Quote:


20051126, 21:39  #18 
Jan 2005
Caught in a sieve
5·79 Posts 
I've been trying to find a generic top5000 prime with LLR 4.62 recently. I chose Proth tests with n=409600, k increasing from 600 to 1000000, since 2^20 is 1048576.
But around k=65487, running on a P4 2.8GHz nonhyperthreaded, I noticed that LLR said I was using a zeropadded FFT. Which according to the thread above shouldn't happen, right? Looking through the log, there seems to have been a big jump in time taken between 60000 and 62000. While I can't completely rule out other processes for this jump, considering the size of the k, that zeropadded message is weird. 
Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
LLR Version 3.8.5 is available!  Jean Penné  Software  11  20110220 18:22 
LLR Version 3.8.1 is now available!  Jean Penné  Software  30  20100921 16:43 
Which version for PIII's?  richs  Software  41  20090107 14:40 
LLR  new version  Cruelty  Riesel Prime Search  8  20060516 15:00 
Version 24.14  Prime95  Software  13  20060215 16:32 