Reserving Cleared Mersenne Number

[danc2]

Hello,

I would like to reserve a previously cleared Mersenne Prime (82589933 to be exact) and run Prime95/CUDALucas with it, but not report it.

The reason I want to do this is to test a new method/program I am using on particular hardware in comparison to Prime95/CUDALucas's speeds on that same hardware.

P.S. I realize I probably am not going to beat either of Mersenne's programs, but I am curious to see how close I can get.

[axn]

Originally Posted by danc2

I would like to reserve a previously cleared Mersenne Prime (82589933 to be exact) and run Prime95/CUDALucas with it, but not report it.

If you're not going to report it, you don't need to reserve it.

[danc2]

*How do I run Prime95/CUDALucas with this Mersenne number?

[Viliam Furik]

Originally Posted by danc2

*How do I run Prime95/CUDALucas with this Mersenne number?

You can go to Advanced -> Test and put in the exponent. That is if you have clear worktodo.txt for Prime95 because I think it puts it on the end of the waiting list. If it does, open worktodo.txt in the folder of Prime95 and copy this line to the first row: "Test=82589933".

If I am wrong, and the Test option immediately starts the requested test, then you don't have to worry about the worktodo.txt.

CUDALucas has the same format for worktodo.txt, IIRC.

[Uncwilly]

Originally Posted by Viliam Furik

If it does, open worktodo.txt in the folder of Prime95 and copy this line to the first row: "Test=82589933".

There is a bit more to it than that.

If mprime or Prime95 is running, stop it (exiting it is best practice). Then open up worktodo.txt and the the line
Test=N/A,82589933,75,1
The N/A tells it and PrimeNet that you are not trying to register it as an assignment. Since it is a known prime, all of the DC work has been done. We don't have to wrry about stepping on anyone's toes.
The ,75,1 tells Prime95 how much trial factoring has been done (and it won't try to do anymore.
Save worktodo.txt and restart Prime95.

[danc2]

Thanks, Villiam and Uncwilly!

I decided to do the smaller exponent 756839 because my program/method is way too slow . If anyone was curious here were the results:

Specs: Intel Xeon @ 2.20GHz, 1 core, 2 threads, 64-bit.

Prime95: ~2 minutes 15 seconds.

My program: ~2 hours 7 minutes.

[Uncwilly]

Try a few scattered non-primes to make sure you get matching residues.

[danc2]

Oh, good idea. My program is not that complex and doesn't use residues (I would love to optimize this program perhaps one day to do so), but I will try out different non-primes as suggested!

[Uncwilly]

If it is or is not prime, if you are using LL, you wind up with a result at the end step.
If it is prime, it is zero. If not, you have a non-zero residue. By comparing 2 runs, that is how we know that the testing produced good results. Someone can tell you how to convert it to the hex format that is used (if you are using decimal).

[kriesel]

Originally Posted by danc2

Oh, good idea. My program is not that complex and doesn't use residues (I would love to optimize this program perhaps one day to do so), but I will try out different non-primes as suggested!

You can greatly speed up correctness testing by adding output of res64 (least significant 64 bits) to your program for selected interim iteration counts. Especially for larger exponents (and fft lengths if you are using that multiplication method).
See https://www.mersenneforum.org/showthread.php?t=24003 which contains tables of selected iteration counts and exponents, for LL and PRP3, to which to compare them.

How does run time scale with exponent in your program?
The fastest programs use Crandall & Fagin's irrational-base discrete weighted transform and take not much more than proportional to exponent squared.
Ordinary long multiplication for squaring will take proportional to exponent cubed; much longer.
Gpuowl run time scaling for PRP3: https://www.mersenneforum.org/showpo...6&postcount=10
Prime95 run time scaling for PRP3, low p: https://www.mersenneforum.org/showpo...78&postcount=2
CUDALucas run time scaling for LL: https://www.mersenneforum.org/showpo...23&postcount=2

[danc2]

Originally Posted by Uncwilly

If it is or is not prime, if you are using LL, you wind up with a result at the end step.
If it is prime, it is zero. If not, you have a non-zero residue. By comparing 2 runs, that is how we know that the testing produced good results. Someone can tell you how to convert it to the hex format that is used (if you are using decimal).

Ahhh, okay. Yes, the residue I was not aware what it was referring to and I was only checking if the S_p-2 == 0. My method, in its infancy at least, is using the gmpy2 library (made in Python, but essentially all c-extensions) and trying to make use of Google Colab's Tensor Processing Unit (TPU) which is faster at multiplications and additions (I haven't read that it is any slower at divisions or subtractions). Consequently, making a hex of the same length as Prime95 is a little odd. If anyone knows how to create one, it would be appreciated!

Thank you for the resources Kriesel! Moving forward, I will use all those residues on any new updates to my progress -- thanks so much for compiling that also!