Prime Gap News

[danaj]

I wrote the script and the Perl module, including the GMP code for sieving, next_prime, and the BPSW test.

There shouldn't be any overlap unless p=11 in the first case. Most of those searches are very low. I have a couple looking for all square-free divisors below 500k but that is only haphazardly running so go for it.

I will likely only be a casual searcher as I have been unable to get the script working with gmp optimizations on windows.

Ouch, really? It would seem that without GMP the programs would run absurdly slow.

For PFGW, you could use just a threshold appropriate for the gap size, then use a second little Perl script to cull the candidates based on actual merit.

PFGW's nextprime is quite slow unless it has been changed since 3.7.7. nextprime(10^2000) is over 2 minutes vs. 13 seconds for my code. It looks like it increments by two from the previous odd, checks divisibility by the first 500 primes using trial division, then calls David Cleaver's strong BPSW function. The latter is called far more often and the actual test is close to 2x slower than my BPSW code.

OTOH, if you're really not running my GMP code, then it would be faster. I'm again surprised it doesn't work with GMP.

perl -Mntheory=:all -E "say prime_get_config->{gmp}"

from the DOS prompt should output 1.

[henryzz]

I couldn't install Math::Prime::Util::GMP into activeperl. It needs gmp installing manually. I will try again getting that working at some point.
Currently I am running on linux but I spend most of my time in windows.

[danaj]

Originally Posted by henryzz

I couldn't install Math::Prime::Util::GMP into activeperl. It needs gmp installing manually. I will try again getting that working at some point.
Currently I am running on linux but I spend most of my time in windows.

Ah, ActiveState. No GMP libraries on their build machines, so they don't build the module. Strawberry Perl for Windows includes the GMP libraries. It's far better for developers than ActiveState (it even includes a C compiler and make so you can install your own packages rather than relying on ActiveState to do it for you).

[henryzz]

Originally Posted by danaj

Ah, ActiveState. No GMP libraries on their build machines, so they don't build the module. Strawberry Perl for Windows includes the GMP libraries. It's far better for developers than ActiveState (it even includes a C compiler and make so you can install your own packages rather than relying on ActiveState to do it for you).

I had just downloaded strawberryperl just before reading this. Installing now. Hopefully this will do everything I need.

[henryzz]

Originally Posted by henryzz

I had just downloaded strawberryperl just before reading this. Installing now. Hopefully this will do everything I need.

Installed and working.

I will stop my Linux runs soon and move them to windows.

[Antonio]

Here's my stats for this weeks submission (checked against merits.txt dated 2015-10-14).
Maintaining >18 records/day/core, despite taking one (of four) core out to try a few search modification ideas.
Nothing spectacular to report, but I live in hope

     Gap   -   Entries 
    Owner  -   Improved
  Rosnthal -      143
  Jacobsen -       83
  RobSmith -       21
  M.Jansen -       17
  Toni_Key -        7
  PierCami -        4
  TorAlmJA -        4
  Andersen -        2
  JLGPardo -        1
  MJPC&JKA -        1
   Gapcoin -        1
     Total -      284
523 new entries for merits.txt
239 first time gaps found
Smallest merit increase        =  0.009355
Largest merit increase         = 11.372329
Largest merit found            = 25.683840
Largest merit (first time gap) = 21.487940
Smallest gap recorded          =     13046
Largest gap recorded           =    219390

[robert44444uk]

Originally Posted by Antonio

Here's my stats for this weeks submission (checked against merits.txt dated 2015-10-14).
Maintaining >18 records/day/core, despite taking one (of four) core out to try a few search modification ideas.
Nothing spectacular to report, but I live in hope

I had to wait a while for anything, but with only 4 cores working on this for 3 months, I have some really pleasing results.

What value are you placing on "my $delta"?

If you have this as zero you are doing twice the work you really need. If you are really interested in merits >20, you should set this to a minimum of 8. You get through the range twice as quickly, but are unlikely to miss too many merits >20

[danaj]

Nice results, and a overall top-20 from Spielauer this week.

I'm still using delta=3. It still gives a nice boost without dropping too many. I am not getting close to that efficiency on most of my searches. Once again contemplating dropping some of the 2*3*5*p / 2*3*5*7*p / 2*3*5*7*11*p ones in favor of something simpler.

[Antonio]

I'm using delta = 2, after some trial runs to determine the effect various deltas would have.

I found that for deltas of 1, 2 and 3 merits the % of gaps with merits >10 that were missed were 6.6%, 9.7% and 18.2% respectively and decided to opt for losing less than 10%.

[robert44444uk]

Originally Posted by Antonio

I'm using delta = 2, after some trial runs to determine the effect various deltas would have.

I found that for deltas of 1, 2 and 3 merits the % of gaps with merits >10 that were missed were 6.6%, 9.7% and 18.2% respectively and decided to opt for losing less than 10%.

That is for merits of 10. So effectively, for gaps of 100k or more - there are too many found <10 that are not records at lower gap targets.

It would be interesting to look at all of your gaps >20 you have found and find what actual prevprime merits were (total merit = prevprime merit+ nextprime merit). My bet is that almost all of the prevprime merits would be >8.

[Antonio]

Originally Posted by robert44444uk

That is for merits of 10. So effectively, for gaps of 100k or more - there are too many found <10 that are not records at lower gap targets.

It would be interesting to look at all of your gaps >20 you have found and find what actual prevprime merits were (total merit = prevprime merit+ nextprime merit). My bet is that almost all of the prevprime merits would be >8.

A casual inspection of my results indicate that this is almost certainly true. When I started I was more interested in filling in the missing gaps rather than finding record merits. I may well move a core over to record merit searching in a week or two, just to see what happens