mersenneforum.org (https://www.mersenneforum.org/index.php)
-   15k Search (https://www.mersenneforum.org/forumdisplay.php?f=16)
-   -   The fastest way to a top-5000 prime? (https://www.mersenneforum.org/showthread.php?t=3871)

 lsoule 2005-03-15 19:53

The fastest way to a top-5000 prime?

I’ve been searching for primes of the form k*2^n-1 with the 15k project for a while now and became interested in finding candidates, k, that would produce a top-5000 prime the fastest. How long it takes to find a top-5000 prime for a particular k value depends on many factors: how efficient sieving is, how long LLR takes for that particular k, and how often that k produces primes. To estimate this, I started with k’s formed by all combinations of the first 11 primes and k>1000. For each k, measure
• The number of candidates left after sieving n=0-10000 to 100M
• The number of primes in the range n=0-10000
• The number of bits in k (the time for LLR is proportional to this)

The result of this can be summed up in a couple of lists. First is the top-20 list of the k’s that had the most primes in the range n=0-10000. All of the k’s here are divisible by 15k.

[CODE]# Primes K
81 8331405
77 944876594805
76 7526103585
73 743411955
72 685084785
71 640049865
71 41237498445
69 1818438765
69 169215941895
69 152125131763605
68 49335
68 965426385
68 1272300315
67 285413005185
67 1212935295
67 553437885
67 51449055
67 377481716535
66 578931045
66 3009765[/CODE]

Next is the top-20 list for the k’s that are estimated to be fastest to find a top-5000 prime with. The score is related to how long it should take to find a top-5000 prime in the range of k*2^200000-1. Thus, the smaller the better. All but 2 of the top 20 are divisible by 15.

[CODE]Score K Primes already on top-5000 list
2481891 8331405 None!
2545731 49335 184364, 199133, 242161
2552143 26565 198349, 217001
2744279 25935 None!
2819593 6555 None!
2850866 19635 189197
2955186 373065 None!
2990424 102765 None!
3013362 62985 None!
3062642 67773 None!
3085355 5865 None!
3098476 3009765 None!
3119285 7526103585 None!
3123287 743411955 None!
3180954 2330445 None!
3197466 19437 None!
3222459 465465 None!
3238416 2805 185593, 192628, 200027, 200212, 203574, 212227, 230666
3242361 685084785 None!
3244999 2667885 None![/CODE]
I tried out the first K, 8331405, and found a top-5000 prime within 24 hours!

 lsoule 2005-03-16 14:50

This one found 5 prime from n=183-221k. Based on the estimates
above, I only expected one prime, but happily surprised:
8331405*2^207555-1
8331405*2^205885-1
8331405*2^199127-1
8331405*2^198122-1
8331405*2^186251-1

 Kosmaj 2005-03-17 02:43

Your approach is interesting. AFAIK when selecting k's nobody tried to take into account the fact that LLR can process small k's much faster than larger ones. Possibly because the first version of LLR released in 2003 couldn't. Also when selecting, or better to say constructing k's (before checking the weight) it was customary (if I'm not mistaken) to select k so that for no n, k*2^n-1 (or +1) is divisible by any small primes (smaller than a certain value) while 8331405*2^n-1 is divisible by 11 and 13 for some values of n.

BTW, do you mind if I try k=25935 from 183,500 to about 210,000.

And congrats on many primes already found for k=8331405.

 lsoule 2005-03-17 03:10

No problem taking 25935. The only other one I'm trying out is 6555.

 Kosmaj 2005-03-17 04:17

OK, thanks!

BTW, some k have been searched before and following primes are recored on Top-5000.
k=25935: 46144, 47396, 117344, 121855, 139225, 152803, 173518
k=373065: 178284, 161917, 150532, 138979

and no other k from the list of 20 fastest marked as primeless.

 Kosmaj 2005-03-17 07:37

I just found out that k=25935 had been already tested by Thomas to 200k. It is listed on the [URL=http://www.15k.org/stats.htm]15k stats page[/URL] but mistakenly as k=29535.

This k is not so promising but I'll test the 200-220k range.

Then I'll switch to k=67773 from 183500 to about 210k.

[Two minutes later]

One of my LLR clients just found that
25935*2^219995-1 is prime!! (66230 digits) :w00t:

It was the very last candidate in the 200-220k range (sieved to 30 bn) and since I started another LLR client to work in the reverse order it was found at once!

 smh 2005-03-17 22:54

The fastest way is probably a fixed N search.

At least thats what it used to be when i was searching for primes. I don't know about changes in sieving, but fixed N sieving used to be much faster. Besides, the size of the number doesn't grow (much) if you don't find a prime.

Just my 2 cents

 lsoule 2005-04-02 21:38

Just for the sake of completeness, I've updated the table this time
searching for all primes with n>100k. My previous list just searched
for primes currently on the top-5000 list...

[CODE]
Score K Primes already on top-5000 list with n>100k
2481891 8331405 186251, 198122, 199127, 200363, 205885, 207555, 220532, 222399, 264459, 269433, 274141
2545731 49335 184364, 199133, 242161
2552143 26565 109141, 112017, 113640, 114935, 124714, 135551, 141212, 145695, 151485, 164728, 179468, 181672, 198349, 217001
2744279 25935 117344, 121855, 139225, 152803, 173518, 219995
2819593 6555 211720, 235260, 236514
2850866 19635 125045, 156800, 189197
2955186 373065 138979, 150532, 161917, 178284
2990424 102765 None!
3013362 62985 None!
3062642 67773 None!
3085355 5865 None!
3098476 3009765 None!
3119285 7526103585 None!
3123287 743411955 None!
3180954 2330445 None!
3197466 19437 None!
3222459 465465 None!
3238416 2805 200212
3242361 685084785 None!
3244999 2667885 None!
[/CODE]

 TTn 2005-04-11 03:07

speed

[QUOTE]The fastest way is probably a fixed N search.
At least thats what it used to be when i was searching for primes. I don't know about changes in sieving, but fixed N sieving used to be much faster. Besides, the size of the number doesn't grow (much) if you don't find a prime.
Just my 2 cents[/QUOTE]

Yes, as far as I know fixed n sieving is much faster. :coffee:
If you use RMA, Newpgen and LLR are automated, and return primes the quickest using this method.

Download RMA from the yahoo primeform group, and then copy MSCOMCTL.ocx, and COMDLG32.ocx, from your windows\system32 folder.
Place them into the same folder as Newpgen, and LLR .

Then click "preferences", "other options", and make sure "RMA enabled", is not checked.

TTn

 lsoule 2005-04-11 17:06

Thanks for the info smh and TTn. I'll take a look at fixed-n sieving.

Now that I've tried out some of the k's in my table, I have another question.
Some k's like 8331405 have 14 primes in the top-5000 from n=198-433k.
Others I have searched n=187-300k and found 0 top-5000 primes. It looks
like the # of primes found in 0-10k isn't always a good predictor of larger primes
(unless the range happens to hit the arbitrarily large gaps between primes...).
Any pointers here?

thanks,
Larry

 TTn 2005-04-12 04:19

[QUOTE]I have another question.
Some k's like 8331405 have 14 primes in the top-5000 from n=198-433k.
Others I have searched n=187-300k and found 0 top-5000 primes. It looks
like the # of primes found in 0-10k isn't always a good predictor of larger primes (unless the range happens to hit the arbitrarily large gaps between primes...).
Any pointers here?[/QUOTE]

Yes, the # of primes, is not always a good predictor of larger candidates.
Nash weight, or Proth weight is also not always a good predictor.

I am including an new option in RMA, that hunts outwards(over under) from the predicted area until a prime is found. Then a new prediction is made based on past data (kn prime) and existing weights, and testing resumes from there. Although a simple idea, it is faily complex to implement, and will skip a small # primes.

Although I should add that we are just making sand castles, as the tide comes in.

All times are UTC. The time now is 21:03.