mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Octoproth Search (https://www.mersenneforum.org/forumdisplay.php?f=63)
-   -   Dodecaproths (https://www.mersenneforum.org/showthread.php?t=5358)

Kosmaj 2006-01-19 04:11

Re: dodeca3
 
[B]R. Gerbicz[/B]
Thanks for your quick reaction! The new version is indeed up to 2 times faster than before. I'll now check whether it correctly detects already known DodecaProths.

Once I wrote a similar program to search for a long arithmetic progression of primes so I have some /limited/ experience.

Greenbank 2006-01-20 12:23

Dodecaproth n 'weights'
 
I'm defining the 'weight' of an n as the number of PRP tests that dodeca_30 needs to perform when checking that n for a range of 0,10T.

The lower the weight the faster it will be to check that n.
The higher the weight the slower it will be.

n=80 weight=45489
n=81 weight=50735
n=82 weight=36134
n=83 weight=31322
n=84 weight=79800
n=85 weight=11925
n=86 weight=71370
n=87 weight=46678
n=88 weight=10297
n=89 weight=452294
n=90 weight=19023
n=91 weight=31650
n=92 weight=57436
n=93 weight=18424
n=94 weight=76983
n=95 weight=17651
n=96 weight=23206
n=97 weight=72811
n=98 weight=77654
n=99 weight=86727

n=89 sticks out somewhat! And I've already taken n=88 :-)

R. Gerbicz 2006-01-20 16:19

[QUOTE=Greenbank]
n=89 weight=452294

n=89 sticks out somewhat! And I've already taken n=88 :-)[/QUOTE]

See my benchmarks for n=89, comparing dodeca_2_5 and 3_0 ( just for [0T,1T] ):
dodeca_2_5:
Number of Prp tests=1225
Time=46 sec.

dodeca_3_0
Number of Prp tests=45391
Time=23 sec.

So it isn't very important the number of prp tests, see my mpz_powm() instruction for n=89 bits modulus takes about say 50000 clock cycles, so if we are checking 45391 prp tests, then that take about 2*10^9 clock cycles or about 1.3 sec. for my pc ( 1.7 GHz ), so the prp testing time is about 5% of the total time, 95% is the sieving time!
That was the reason why I've decreased all parameters in 2_5 version of dodeca program.
I've also defined the weight of n ( for dodeca numbers ). See post #28 in this thread.

Kosmaj 2006-01-24 22:01

R.Gerbicz
 
Why is [I]k[/I] maxed to 2^60 in dodeca.c? There are 4 more bits in 8-byte unsigned int and furthermore, if you keep k/105 instead of [I]k[/I], the range can be extended to 105*(2^64-1) without using gmp variables, not so?

R. Gerbicz 2006-01-24 22:02

new dodeca program version 4.0
 
1 Attachment(s)
In this version there is no limitation on k, and I've also decreased prime_limit from 2000 to 1000 ( and try also -O1 optimization flag ).
See the attachment for the code. Or download exe from [URL="http://www.robertgerbicz.tar.hu/dodeca_4_0.exe"]http://www.robertgerbicz.tar.hu/dodeca_4_0.exe[/URL]

Kosmaj 2006-01-25 14:27

My first tests indicate that ver.4 is about 30% slower than ver.3. Please tell me, can I use ver.3 to test n=103 for k<2^60 (by removing the n>=100 condition from the source).

R. Gerbicz 2006-01-25 14:49

[QUOTE=Kosmaj]My first tests indicate that ver.4 is about 30% slower than ver.3. Please tell me, can I use ver.3 to test n=103 for k<2^60 (by removing the n>=100 condition from the source).[/QUOTE]

In some cases it is possible because it is using sieve_limit=1000. Yes you can change ver.3, by removing that condition.

Note that in all of the new programs now you can change sieve_limit in the source and to do this you have to change also number_of_primes, because there is an error checking for this in the program, you can also change block_length and magic_constant, and of course magic_constant>=block_length>=sieve_limit has to be true . But note that in hexadeca program you can change only sieve_limit ( and number_of_primes ), don't change block_length otherwise the program won't search for the whole interval for n=76 ( there is no magic_constant for this hexadeca ).

I've changed the source in dodeca_4_0 to use sieve_limit=2000 and number_of_primes=303 but this was still slower by 7% for n=99. I don't know why perhaps because we are using in that part mpz_t integers and not long variables.


All times are UTC. The time now is 16:46.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.