Thread: Pari/GP to PFGW
View Single Post
Old 2017-12-21, 01:50   #4
"Dana Jacobsen"
Feb 2011
Bangkok, TH

22·227 Posts

Admittedly the BPSW test should exit early for the composites. Still, PFGW will be faster on the Fermat / M-R once the result hits somewhere between 2000 and 6000 digits -- at least that's where it's faster than GMP's mpz_powm.

As an aside, I have an example Fibonacci prime finder in my Perl module that has a couple parallel implementations. Starting at the beginning searching for each, my timing results are

26244 seconds for Fp36 (p = 148091) on a 3930K (12 proc)
12000 seconds for Fp36 (p = 148091) on a c4.8xlarge (36 proc)
6465 seconds for Fp36 (p = 148091) on a r4.16xlarge (64 proc)

Someday I'll get a big c5 for a day to benchmark.

[Edit: that it not the timing for one test, but the time from starting search for all Fp until we see the output of the in-order 36th Fibonacci prime]

I've tried it on a Power8 with 152 logical processors but it's a shared machine so really couldn't let it run very long, and it looks like GMP is better optimized for x86-64.

It's pretty simple, just running primality tests in order on each core, with the most interesting part just being making sure the results are shown in order.

Last fiddled with by danaj on 2017-12-21 at 01:51
danaj is offline   Reply With Quote