n=1.7M and n=3.322M Twin-Sophie record attempts

[VBCurtis]

When you hold back the big primes, you get the big smileys?

[MooMoo2]

Originally Posted by VBCurtis

When you hold back the big primes, you get the big smileys?

Hmmm...

Anyway, I've done some additional sieving on the n=1.7M file and sieved it to p=1047T (the previous uploaded file was sieved to p=540T). We're at k=178G now, so I've uploaded k=200G - 20T. The file contains 5,709,024 candidates: https://pixeldrain.com/u/LDBtrkJa

Gary, could you replace the k>200G candidates in the server with the candidates in that link? I swear I didn't hold back any primes this time...

[gd_barnes]

Originally Posted by MooMoo2

Hmmm...

Anyway, I've done some additional sieving on the n=1.7M file and sieved it to p=1047T (the previous uploaded file was sieved to p=540T). We're at k=178G now, so I've uploaded k=200G - 20T. The file contains 5,709,024 candidates: https://pixeldrain.com/u/LDBtrkJa

Gary, could you replace the k>200G candidates in the server with the candidates in that link? I swear I didn't hold back any primes this time...

I don't know. I think you're removing the rest of the primes! :-)

This is done.

Balancing n=1.7M for future reference:
k<200G: 62,568 candidates sieved to 540T
k=200G-500G: 93,562 candidates sieved to 540T removed from port 12000
k=200G-500G: 86,602 candidates sieved to 1047T added to port 12000

Currently 62,568+86,602=149,170 candidates in port 12000 for k<500G.

[MooMoo2]

Originally Posted by gd_barnes

I don't know. I think you're removing the rest of the primes! :-)

This is done.

Balancing n=1.7M for future reference:
k<200G: 62,568 candidates sieved to 540T
k=200G-500G: 93,562 candidates sieved to 540T removed from port 12000
k=200G-500G: 86,602 candidates sieved to 1047T added to port 12000

Currently 62,568+86,602=149,170 candidates in port 12000 for k<500G.

Looks good, thanks!

[gd_barnes]

Michael,

What sieving software are you using for this? I'm curious to know how you run the quad sieve for twins and Sophies.

Gary

[Cybertronic]

Originally Posted by gd_barnes

Michael,

What sieving software are you using for this? I'm curious to know how you run the quad sieve for twins and Sophies.

Gary

NewPgen have this feature. [select Twin/SG...]

[MooMoo2]

Originally Posted by gd_barnes

Michael,

What sieving software are you using for this? I'm curious to know how you run the quad sieve for twins and Sophies.

Gary

I checked the "Lucky Minus" option on NewPGen, clicked "verify results", and let it run until p=300T or so.

After that, I fed it into TwinGenX to take advantage of the multithreading. IIRC, the header had to be changed a bit since NewPGen and TwinGenX used slightly different formats.

http://www.underbakke.com/primes/

[gd_barnes]

The prime gods have been unkind to TPS so far for n=1.7M:

k<200G: 62,568 tests, sieve depth P=540T, expected # of primes: 3.21
k>200G so far: 12,079 tests, sieve depth P=1047T, expected # of primes: 0.63

Total: 74,647 tests
Expected # of primes: 3.84
Chance of prime: 97.85%

Actual # of primes: 0

:-(

[MooMoo2]

Originally Posted by gd_barnes

The prime gods have been unkind to TPS so far for n=1.7M:

k<200G: 62,568 tests, sieve depth P=540T, expected # of primes: 3.21
k>200G so far: 12,079 tests, sieve depth P=1047T, expected # of primes: 0.63

Total: 74,647 tests
Expected # of primes: 3.84
Chance of prime: 97.85%

Actual # of primes: 0

:-(

Ouch! Perhaps we need to appease them

On the next full moon, take your fastest CPU to the top of the highest mountain in your area just after sunset and light 7 candles around it. Take a deep breath, say "You are the chosen one to bring glory to TPS", and blow the candles out. The ritual should work

If it doesn't, I wonder if there's an issue with either NewPGen, LLR, and/or PRPNet. Perhaps there's a bug in the code that only shows up when testing high k and n values.

[gd_barnes]

Originally Posted by MooMoo2

Ouch! Perhaps we need to appease them

On the next full moon, take your fastest CPU to the top of the highest mountain in your area just after sunset and light 7 candles around it. Take a deep breath, say "You are the chosen one to bring glory to TPS", and blow the candles out. The ritual should work

If it doesn't, I wonder if there's an issue with either NewPGen, LLR, and/or PRPNet. Perhaps there's a bug in the code that only shows up when testing high k and n values.

That one got me good! Pretty hard to find a mountain smack dab in the middle of the U.S. :-)

NewPGen has been around for ~15-20 years and has proven to be very accurate as I did my very long "all twin" prime search here for all k<1e6 and n<=50000. It may be clunky with today's modern processors but it's super reliable. I'm confident about PRPnet also. I'm only slightly less confident about LLR due to a few issues here and there lately.

I don't know anything about TwinGenX. What has been your experience with it?

I'm not too concerned yet but will start to become so if we pass 100,000 tests without a prime. At 100,000 tests, the chance of prime are > 99.4%.

Just for future reference: Did you happen to keep the files after you were done sieving with NewPGen before you fed them to TwinGenX?

[MooMoo2]

Originally Posted by gd_barnes

That one got me good! Pretty hard to find a mountain smack dab in the middle of the U.S. :-)

NewPGen has been around for ~15-20 years and has proven to be very accurate as I did my very long "all twin" prime search here for all k<1e6 and n<=50000. It may be clunky with today's modern processors but it's super reliable. I'm confident about PRPnet also. I'm only slightly less confident about LLR due to a few issues here and there lately.

I don't know anything about TwinGenX. What has been your experience with it?

I'm not too concerned yet but will start to become so if we pass 100,000 tests without a prime. At 100,000 tests, the chance of prime are > 99.4%.

Just for future reference: Did you happen to keep the files after you were done sieving with NewPGen before you fed them to TwinGenX?

Aside from n=1.7M and n=3.322M, I've never used TwinGenX, so my experience is fairly limited. I do know that PrimeGrid used TwinGen in their n=1.29M search (https://www.primegrid.com/download/S...17_1290000.pdf), but it's unclear as to whether it was done with the original TwinGen released in October 2002 (see: http://www.underbakke.com/primes/), or the newer TwinGenX that was released in May 2012.

In any case, I only used TwinGenX for the higher p values, above p=~300T. The number of candidates closely matched the expected density, so I'm not too concerned. Since the program can't remove primes while leaving composites in the file (this would require TwinGenX to determine primality, which is impossible), the worst-case scenario would be that the program simply removes candidates randomly regardless of whether or not they're prime. In that case, our sieve depth should still be p=~300T with ~3.7 primes expected to be found so far (~97% chance of at least one prime).

After uploading the files to pixeldrain, I deleted the progress files that had the candidates before they were fed to TwinGenX. If we still can't find any primes after ~100,000 tests, the next step should probably be randomly checking low-p ranges on the sieve file (p=10T-11T, etc.) on different hardware with both NewPGen and TwinGenX to see if any candidates are removed. Since it's unlikely that two different programs on different hardware would have the same issue, that would rule out sieving problems and hardware problems.

As for now, my very rough guess for the lack of primes is:
80% chance: the prime gods hate us
10% chance: there's a problem with PRPNet (there are known bugs in using PRPNet to search for Sophies, so we'll be manually testing for them instead. There may be similar issues with high-k values.)
7% chance: there's a problem with LLR
2% chance: there's a problem with the sieve software
1% chance: there's a problem with my hardware and/or the hardware used to LLR n=1.7M (none of my PCs are overclocked, and they've been prime-stable for years. An issue with everyone else's hardware is also very unlikely since there were no problems with the variable-n range, and most testers are the same for both subprojects)