New LLT program in the pipe..

[Dr Autonomy]

Just finished a new Windows-based LLT program that can perform a Lucas-Lehmer Test, with viewable proof of work, on any Mersenne number whose defining binary exponent is between 67 and 4,294,967,291, inclusive.

Which means that it can be used to discover the first one billion decimal digit prime number, as long as the hardware is up to the task. Practically speaking, that probably means a machine with at least 32 processor cores, probably more, but testing hasn't gone that far yet, so I don't actually know..

At this point, I'm just looking for feedback. You can download it here. Just click on "Download" in the top left corner.

The download itself is a zip file with a readable (but very short) "File_ID.diz" file inside.

The operation of the program is very intuitive, so no manual is supplied, other than the "File_ID.diz" file.

Internally, it uses 64K-pt FFTs to do all of its heavy lifting, which was hand-crafted in Intel assembly. However, source code is not available at this time.

If interested, feel free to download and play with it, and let me know what you think!

[retina]

Originally Posted by Dr Autonomy

At this point, I'm just looking for feedback. You can download it here. Just click on "Download" in the top left corner.

Requires JS to do anything, otherwise the page is blank. Plus the risk of link rot etc.

Attaching it here might be a better option.

[Dr Autonomy]

It's a dropbox link, so I don't see how Javascript is involved, but whatever..

So how do I "attach it here"? It's a 90 Meg zip file.

[retina]

Originally Posted by Dr Autonomy

It's a dropbox link, so I don't see how Javascript is involved, but whatever..

The dropbox page is blank without JS. No links, no anything.

[Batalov]

Originally Posted by Dr Autonomy

It's a dropbox link, so I don't see how Javascript is involved, but whatever..

So how do I "attach it here"? It's a 90 Meg zip file.

The source is 90 Meg zip file?
We are not interested in any blind .exe files that would simply mine for bitcoins for your profit for all we know.
Just publish the source.

Can you prove that your executable is benign?

I attached the File_ID.diz file for the benefit of the forum readers. It does not inspire confidence. In fact, it looks more like the second making of the previous very similar posting some time ago. That one also demanded "a significantly powerful machine, with at least 32 Intel-compatible processor cores, or possibly more, to run this program for that purpose."

[kriesel]

Originally Posted by Batalov

it looks more like the second making of the previous very similar posting some time ago. That one also demanded "a significantly powerful machine, with at least 32 Intel-compatible processor cores, or possibly more, to run this program for that purpose."

Link, or forum username?

[kriesel]

Some posts in this thread may seem negative or antagonistic or distrusting.
We've been down this road before, and if it seems to be too good to be true, it generally is.
We would welcome any genuine number theory software relevant to finding Mersenne primes or detecting composites, that is reasonable speed and reliability, but it and its author need to be vetted before we trust our systems to it.
We've seen some cranks, trolls and con artists from time to time, and some excellent programmers and fast good legit code. It takes time, but we can tell the difference.

The dropbox page lists the following:
File_ID.diz 1.65 KB
GP_LLT.exe 209 KB
MC.db 92.91 MB
Proof.exe 149 KB
VC_redist.x64.exe 24.29 MB

First reaction:
No source code = low acceptance if any at all. There are very few people who can get us to run code without source. (His name is George Woltman. Not someone hiding behind an alias with no track record here.)
The leading software packages for all hardware and OS and computation type choices are entirely open source or nearly entirely.
And that emphatically includes the work of world-class programmers producing performance faster than the major hardware manufacturers' library routines on the same hardware.
There will be no test drive reactions until someone creates an adequately secured VM sandbox in which to experiment, and spends some time doing so.

Meanwhile, questions abound:

1 What is the purpose & content of the large MC.db?
2 What mechanism is used for proof of work of an LL test?
3 How does a 64k point fft perform work that requires up to 256 M DP word fft length in Mlucas?
4 What testing have you done to determine run time scaling? (on whatever n-core hardware you have available)
5 What testing have you done to prove that it performs iterations accurately, producing correct interim full length residues?
6 What Intel processor features are supported/required by the assembly code? (I have examples of 386, 486, pentium, ..., i7-4790, i5-1035g1, i5-1165g7, dual xeon, xeon phi, etc., most of which are in running condition. I think there's even an NEC 70108 around somewhere.)
7 What version(s) of Windows should it run on or does it require? Vista, 7, 8.x 10, 11? Specific updates (21H2 etc?) Is 64 bit required?
8 Is there a Linux version available? Distros, versions, bitness supported/required?
9 There seems to be little or no included documentation. How is either program used? How does one specify which exponent to test? How to store a computation in progress before closing the program, resume later, etc?
10 Does it support some standard worktodo.txt input format, or results output format?
11 Why choose LL rather than the PRP3 with far superior error checking possibilities & proof generation & certification?
12 How do you get a reasonable amount of function in a mere 209 KB, in stark contrast to the sizes of Mlucas, mprime, gpuowl, CUDALucas etc.?
13 Would you be willing to run tests on at least two exponents we select, well spaced on a log scale, on hardware you provide and describe in detail, so that we can gauge the run time and scaling? With documentation & disclosure of run time and GP_LLT.exe and proof.exe output for them?
14 What multiplication algorithm does it use?
15 What error detection/correction mechanisms does it use?

[Dr Autonomy]

Most of the bulk of the zip is a "Mersenne Composite database" (MS.db), whose sole purpose is to enable the exclusion of known Mersenne composites from the defining binary exponent selection process. If not found, GP_LLT will simply not exclude any qualified Mersenne prime candidates from the selection process, regardless of whether or not they have been factored, for example.

So here's the abridged zip file:

[Dr Autonomy]

Wow. You type fast. I'll attempt to address your numbered points:

1) See my last post.
2) A binary file is produced that consists of one record per iteration of the LLT test. The iteration number is implied, and each record consists of the 64-bit Residue, followed by the time, in milliseconds, that the iteration took to complete.
3) The 64K-pt blocks (32,768 Dwords) are the digits of the system, and for a billion decimal digit number, there are therefore 4,096 digits to deal with. All digit multiplications are done in parallel.
4) Testing has been severely limited by the fact that I only own an Intel I-5. However, this program has never been wrong, in my experience. I've successfully proven the primality of M6972593, a 2,098,960 decimal digit candidate on my puny machine (two of the cores are disabled).
5) See 4..
6) It requires a "modern" 64-processor, and the program itself will check the processor's capabilities, and error out (gracefully) if your processor is not 'up to snuff'..
7) To be fair, it's only been tested on Win 10. I've tried my best to program it to be compatible with Win 7 and above, but that's never been tested..
8) No Linux version.
9) The program itself walks you through the interface as it executes, as far as humanly possible.
10) No. Once you select the defining binary exponent, it just begins the LLT, at which point you click the "Start" button, and off you go. However, while that is running, you can invoke the same program again, to perform either another LLT, or a pre-LLT trial factoring attempt (which is an option).
11) Not looking for a 'probable prime'. Looking for Proof of Primality.
12) It doesn't use NVidia, for starters. Only Intel. Fed up with NVidia. Too many restrictions.
13) Run time and scaling are among the things I'm hoping to find out more about by getting feedback..
14) Grade school squaring, in parallel.
15) CRC32 for file integrity.

Hope that helps..

[retina]

Originally Posted by Dr Autonomy

14) Grade school squaring, in parallel.

That is very bad. It doesn't matter how much "parallel" you have, the algorithm is awful.

It will be many times slower than what we use now.

[kriesel]

Originally Posted by Dr Autonomy

2) A binary file is produced that consists of one record per iteration of the LLT test. The iteration number is implied, and each record consists of the 64-bit Residue, followed by the time, in milliseconds, that the iteration took to complete.
3) The 64K-pt blocks (32,768 Dwords) are the digits of the system, and for a billion decimal digit number, there are therefore 4,096 digits to deal with. All digit multiplications are done in parallel.
4) Testing has been severely limited by the fact that I only own an Intel I-5. However, this program has never been wrong, in my experience. I've successfully proven the primality of M6972593, a 2,098,960 decimal digit candidate on my puny machine (two of the cores are disabled).
5) See 4..
6) It requires a "modern" 64-processor, and the program itself will check the processor's capabilities, and error out (gracefully) if your processor is not 'up to snuff'..
7) To be fair, it's only been tested on Win 10. I've tried my best to program it to be compatible with Win 7 and above, but that's never been tested..
8) No Linux version.
9) The program itself walks you through the interface as it executes, as far as humanly possible.
10) No. Once you select the defining binary exponent, it just begins the LLT, at which point you click the "Start" button, and off you go. However, while that is running, you can invoke the same program again, to perform either another LLT, or a pre-LLT trial factoring attempt (which is an option).
11) Not looking for a 'probable prime'. Looking for Proof of Primality.
12) It doesn't use NVidia, for starters. Only Intel. Fed up with NVidia. Too many restrictions.
13) Run time and scaling are among the things I'm hoping to find out more about by getting feedback..
14) Grade school squaring, in parallel.
15) CRC32 for file integrity.

Hope that helps..

Thanks. Some follow-ups, if you'll indulge me further.
For 2), what is a typical time in milliseconds on the M6972593 you ran, on how many i5 cores (and what full model number of that cpu is that; as in, what Intel cpu model number or part number)?
4) How do you know the program has never been wrong? It's common that programs in this area can produce correct results usually and error sometimes. Software bugs are like that. Hardware reliability is like that. Certain inherent algorithmic errors are like that. Software development is like that.
How long did the M6972593 run take overall, start to finish, of actual run time?
6) What Intel instruction set extensions are required, or sufficient, by standard name?
9) Some screen shots please.
10) Tell me about your TF algorithm.
11) PRP is accepted as the superior standard, because of far superior error detection, and has been for years. It returns definitely composite, at extremely high reliability, far more reliable than LL. And almost everything we test turns out to be composite, by either test. If a PRP test returns PRP-P instead of PRP-C, a retest in parallel using multiple software programs on multiple hardware types by multiple people working independently is the standard verification process, and no one begrudges that effort, since the lower error rate of PRP in practice more than pays for it. If you know of a way to make LL as reliable as PRP in practice, please enlighten us how. With references to the number theory papers.
12) Neither Mlucas nor mprime use Nvidia. Please answer the question.
13) Again, would YOU be willing to run timed tests, on YOUR software, on YOUR hardware, on exponents WE select? YOU're the one asking us to run YOUR software without disclosing ANY SOURCE CODE.
14) Your reply does not bode well for run time scaling. O(p³) Even though the proportionality constant may be low compared to full-grammar-school by use of the lower level 64k point fft. Mlucas mprime etc are fft all the way, and multithreaded, and hand tuned assembly, and OBD PRP/LL is not very feasible in human & hardware lifetime scales. "Mere" gigaBIT tests require patience for a few months' run on fast GPUs.
15) What other error control mechanisms?
16) "hand-crafted in Intel assembly" for which processor models/instruction sets/cache characteristics?