FastECPP software and 50000 digit primality proof (reposted from NMBRTHRY)
[QUOTE="Andreas Enge @inria.fr"]
Dear colleagues, I am pleased to announce the release of a new free software for creating elliptic curve primality proofs using the fastECPP algorithm due to Morain, Franke, Kleinjung and Wirth [1, 2]. It is available under the GPL version 3 or later at [url]https://www.multiprecision.org/cm/[/url] It relies on the approach of computing class polynomials by complex approximations as described in [3]. Optimal class invariants are chosen derived from Weber functions [4], simple [5] or double eta quotients [6], including cases where it is enough to compute lowerdegree subfields of the class field [7]. The evaluation of modular functions, which is the most important part of the class polynomial computation, is optimised following [3, 8]. To ease the step of factoring class polynomials modulo primes, the class fields are then represented as a tower of cyclic Galois extensions of prime degree following [9]. The software relies on a number of libraries from the GNU project, notably GMP [10], MPFR [11] and MPC [12], and on PARI/GP [13] for computations with class groups and for root finding modulo a prime. The code has been used to prove a record prime of about 50000 digits; indeed it has shown that [URL="http://factordb.com/index.php?id=1100000000213115214"]10^50000+65859[/URL] is the smallest prime with 50001 digits. The certificate is available at [url]https://www.multiprecision.org/downloads/ecpp/cert50000.bz2[/url] in PARI/GP format [13] and at [url]https://www.multiprecision.org/downloads/ecpp/cert50000.primo.bz2[/url] in Primo format [14] as converted by PARI/GP code written by J. Asuncion, who is also the author of the fastECPP implementation in PARI/GP. Parallelisation uses MPI. Computations have been carried out on the PlaFRIM [15] and MCIA [16] clusters in Bordeaux. The first phase of the record, in which a list of discriminants and candidate orders of points on elliptic curves is produced, has taken about 26 days of wallclock time and 67 years of CPU time, in several runs on clusters with 752 to 1328 cores. Of this CPU time, about 8% has been devoted to the computation of square roots modulo a prime by the TonelliShanks algorithm, about 12% to solving quadratic equations by Cornacchia's algorithm, about 65% to removing smooth factors from curve cardinalities, and about 15% to internal primality tests. This imbalance probably indicates nonoptimal choices of parameters. On the other hand, it leads to very short certificates: We reach a length of only 1645 steps, as opposed to the 3794 steps in the recent 49000 digits record [17], which implies that verification time is proportionally lower. For the specialists, the parameters were as follows: Discriminants with absolute value up to about 6.9*10^9, leading to class numbers with prime factors up to 157 (and no bound on the class numbers themselves), and a smoothness bound of about 5.5*10^11 for partial factorisation of point orders. The second phase of the record, in which complex multiplication produces a list of elliptic curves and points of (recursively shown to be) prime orders on them, has taken about 74 days of wallclock time and less than 4 years of CPU time on a few machines with 32 to 256 cores in total. Depending on the largest factor of the class number, each step may take a vastly differing amount of time. The longest step has taken close to 10 days for factoring a class polynomial for the discriminant 345992327 of class number 13112=2^3*11*149 for an intermediate prime of about 47800 digits; this would also have been the wallclock time for this phase had it been run sufficiently in parallel. Of the total CPU time, about 96% has been devoted to finding roots of class polynomials, about 3% to verifying point orders, and only about 1% to the construction of the class polynomials. Verification of the certificate takes a little over 4 hours using PARI/GP on a machine with 128 cores. Andreas Enge [1] F. Morain: "Implementing the asymptotically fast version of the elliptic curve primality proving algorithm", Mathematics of Computation 76 (257), 2007, pp. 493505 [2] J. Franke, T. Kleinjung, F. Morain and T. Wirth: "Proving the Primality of Very Large Numbers with fastECPP", in Duncan Buell: "Algorithmic Number Theory  ANTSVI", Lecture Notes in Computer Science 3076, SpringerVerlag, Berlin 2004, pp. 194207 [3] A. Enge: "The complexity of class polynomial computation via floating point approximations", Mathematics of Computation 78 (266), 2009, pp. 10891107 [4] R. Schertz: "Die singulären Werte der Weberschen Funktionen $f$, $f_1$, $f_2$, $\gamma_2$, $\gamma_3$", Journal für die reine und angewandte Mathematik 286/287, 1976, pp. 4674 [5] A. Enge and F. Morain: "Generalised Weber Functions", Acta Arithmetica 164 (4), 2014, pp. 309341 [6] A. Enge and R. Schertz: "Constructing elliptic curves over finite fields using double etaquotients", Journal de Théorie des Nombres de Bordeaux 16, 2004, pp. 555568 [7] A. Enge and R. Schertz: "Singular values of multiple etaquotients for ramified primes", LMS Journal of Computation and Mathematics 16, 2013, 407418 [8] A. Enge, W. Hart and F. Johansson: "Short addition sequences for theta functions", Journal of Integer Sequences 18 (2), 2018, pp. 134 [9] A. Enge and F. Morain: "Fast Decomposition of Polynomials with Known Galois Group", in Marc Fossorier, Tom Høholdt and Alain Poli (editors): "Applied Algebra, Algebraic Algorithms and ErrorCorrecting Codes  AAECC15", Lecture Notes in Computer Science 2643, SpringerVerlag, Berlin 2003, pp. 254264 [10] Torbjörn Granlund et al.: "GMP  The GNU Multiple Precision Arithmetic Library", release 6.2.1, 2020, [url]http://gmplib.org/[/url] [11] Guillaume Hanrot, Vincent Lefèvre, Patrick Pélissier, Paul Zimmermann et al.: "GNU MPFR  A library for multipleprecision floatingpoint computations with exact rounding", release 4.1.0, 2020, [url]http://www.mpfr.org/[/url] [12] A. Enge, M. Gastineau, P. Théveny and P. Zimmermann: "GNU MPC  A library for multiprecision complex arithmetic with exact rounding", release 1.2.1, 2020, [url]https://www.multiprecision.org/mpc/[/url] [13] PARI Group: "PARI/GP", release 2.13.4, 2022, [url]https://pari.math.ubordeaux.fr/[/url] [14] M. Martin: "Primo", release 4.3.3, 2020, [url]http://www.ellipsa.eu/public/primo/primo.html[/url] [15] PlaFRIM, Plateforme Fédérative pour la Recherche en Informatique et Mathématiques, [url]https://www.plafrim.fr/[/url] [16] MCIA, Mésocentre de Calcul Intensif Aquitain, [url]https://www.mcia.fr/[/url] [17] [url]https://primes.utm.edu/primes/page.php?id=133761[/url] [/QUOTE] Exciting development. The early preview was visible since last August: [QUOTE]rank prime digits who when comment 64129 U(148091) 30949 x49 Sep 2021 Fibonacci number, ECPP 66197 (2^95369 + 1)/3 28709 x49 Aug 2021 Generalized Lucas number, Wagstaff, ECPP[/QUOTE] 
An amazing feat! Such parallelism and clever mathematics run on some very big hardware. I take my hat off to these provers. :smile:

So 71 years of CPU time for 51,000 digits vs 20 months * 64 cores ~ 107 years of CPU time for 49,081 digits.
But depends on how comparable those CPU core years are. 51,000 digits is supposed to take (51000/49081)^3.75 ~ 15% longer ? At least with Primo. 
Note in their text that they say "This <<...>> probably indicates nonoptimal choices of parameters. On the other hand, it leads to very short certificates: We reach a length of only 1645 steps" (!!).
They call it a prerelease. They will likely optimize it even more. Btw, I had to "massage" their cert a little bit, but then factorDB took it. Took 3 iterations slightly reformatting the file (and to rename it into .out file) 
[QUOTE=ATH;605381]So 71 years of CPU time for 51,000 digits vs 20 months * 64 cores ~ 107 years of CPU time for 49,081 digits.
But depends on how comparable those CPU core years are. 51,000 digits is supposed to take (51000/49081)^3.75 ~ 15% longer ? At least with Primo.[/QUOTE] It is 50,001 digits. My estimate for Primo is (50001/49081)^4 = 1.077 or 7.7% longer. I used quad channel DDR 2400MHz for R49081's certification. I guess they were using octa channel most of the time, but that does not explain their speed fully. There is marginal gain by having fewer stage 2 steps to do. 
[QUOTE=Batalov;605382]Note in their text that they say "This <<...>> probably indicates nonoptimal choices of parameters. On the other hand, it leads to very short certificates: We reach a length of only 1645 steps" (!!).
They call it a prerelease. They will likely optimize it even more. Btw, I had to "massage" their cert a little bit, but then factorDB took it. Took 3 iterations slightly reformatting the file (and to rename it into .out file)[/QUOTE] Please email Andraes and ask him to make an UTM submission. 
I made a shorter jump and forwarded NMBRTHRY email to Caldwell. (though he is surely on the list.*)
________ * Funny story: a few years ago a ton (literally!) of porn links started coming every day via NMBRTHRY email list. It went on for quite a while. Many folks, I am afraid, unsubscribed from it at that point. Since then it has been moderated with huge delays. One curious example was when an invitation to a December conference was sent out in March. 
This is great news for two reasons:
1. It's a new ECPP world record. 2. We finally have an opensource ECPP implementation that can compete with Primo. I'm aware that Marcel has no obligation to make Primo open source, but I do feel that closedsource software goes against the spirit of science. 
Holy s**t!
Hats off and congrats to all those involved. That is some serious soft and hardware. I wonder how big the code can go? There are so many interesting numbers from 50K to 100K digits. Just the electricity cost alone will make such a number pretty valuable :smile: 
Fantastic!
It would seem [URL="https://primes.utm.edu/top20/page.php?id=27"]this page[/URL] needs an update? 
Some entity should set up fastECPP for a proofoverthenet. Clients could latch onto the main server. This way a 60k proof might be computed in a reasonable amount of time.

All times are UTC. The time now is 09:00. 
Powered by vBulletin® Version 3.8.11
Copyright ©2000  2022, Jelsoft Enterprises Ltd.