Pépin tests of Fermat numbers beyond F24

[ET_]

Too bad I still don't owe a Haswell...

Luigi

[LaurV]

Originally Posted by ET_

Too bad I still don't owe a Haswell...

Luigi

Hm, do you want to owe it, or to own it?

[Xyzzy]

Originally Posted by ewmayer

…although should you change your mind (or blow up your system) at some point along the way you can hand things off (even if your savefiles get nuked, since we will be cross-checking those vs mine) to a fresh recruit.

[ewmayer]

Originally Posted by ET_

Too bad I still don't owe a Haswell...

Luigi

Now you have the perfect excuse to run out and buy one! You can probably get used systems/components fairly cheap now.

Re. ramshanker's "I envy these algorithms" comment -- you envy abject #FAILure?

[ET_]

Originally Posted by ewmayer

Now you have the perfect excuse to run out and buy one! You can probably get used systems/components fairly cheap now.

I will, probably for my birthday on March 7th. Will I be able to help then, or will it be too late?

[ewmayer]

Originally Posted by ET_

I will, probably for my birthday on March 7th. Will I be able to help then, or will it be too late?

Not at all!

[ET_]

Originally Posted by ewmayer

Not at all!

Fine, let's keep in touch then

[ewmayer]

Originally Posted by ET_

Fine, let's keep in touch then

While my 32M Haswell run continues - at ~900,000 iterations per day - I'll be using spare cycles on my slow Core2 Macbook to re-run the roughly 1 million iterations from Mike's last 32M savefile (whose last checkpoint data agreed with those of my just-completed 30M run at the iteration in question) to the point at which my 30M and 32M-FFT runs diverge, in order to ascertain which of the 2 runs went off the rails. (Like I said I'm almost certain it was the 30M one, but I need to confirm that via a 3rd run.)

[ewmayer]

Originally Posted by ewmayer

While my 32M Haswell run continues - at ~900,000 iterations per day - I'll be using spare cycles on my slow Core2 Macbook to re-run the roughly 1 million iterations from Mike's last 32M savefile (whose last checkpoint data agreed with those of my just-completed 30M run at the iteration in question) to the point at which my 30M and 32M-FFT runs diverge, in order to ascertain which of the 2 runs went off the rails. (Like I said I'm almost certain it was the 30M one, but I need to confirm that via a 3rd run.)

It's official - re-run of iters 24.92M (Where Mike left off his Haswell-NUC DC run @FFTlen 32M) - 25.64M (where my Haswell-quad continuation of same first mismatches my original run @FFTlen 30M) using the above-mentioned SSE2 build on my Core2Duo Macbook
@FFTlen 30M matches my (ongoing) Haswell 32M continuation run:

[1] Original Haswell-quad run @30M:

[Jun 30 12:51:11] F29 Iter# = 25630000 clocks = 00:17:59.766 [ 0.1080 sec/iter] Res64: C782D1DA346F61C4. AvgMaxErr = 0.118563074. MaxErr = 0.156250000
[Jun 30 13:09:15] F29 Iter# = 25640000 clocks = 00:18:00.266 [ 0.1080 sec/iter] Res64: BCF83642C5E19A09. AvgMaxErr = 0.118375359. MaxErr = 0.148437500

[2] DC Haswell-NUC/quad run @32M:

[Jan 21 14:55:02] F29 Iter# = 25630000 clocks = 00:15:58.792 [ 0.0959 sec/iter] Res64: C782D1DA346F61C4. AvgMaxErr = 0.023437500. MaxErr = 0.023437500
[Jan 21 15:11:06] F29 Iter# = 25640000 clocks = 00:15:59.235 [ 0.0959 sec/iter] Res64: 12AF4B8DD106DF85. AvgMaxErr = 0.023437500. MaxErr = 0.023437500

[3] DC Core2Duo/SSE2-build run @30M (timings vary widely since only run when plugged in):

[Feb 03 23:55:27] F29 Iter# = 25630000 [ 4.77% complete] clocks = 02:57:20.963 [ 1.0641 sec/iter] Res64: C782D1DA346F61C4. AvgMaxErr = 0.102656999. MaxErr = 0.132812500.
[Feb 04 18:19:56] F29 Iter# = 25640000 [ 4.78% complete] clocks = 05:04:33.375 [ 1.8273 sec/iter] Res64: 12AF4B8DD106DF85. AvgMaxErr = 0.102699611. MaxErr = 0.140625000.


Thus, everything below beyond iter 25.63M in my original Haswell/AVX2 run @FFTlen 30M is garbage.

Luigi, you order your birthday present yet? My hope is that things will run sufficiently fast on your newer hardware that you might get throughput matching my 4-core speed using just 3 (or even 2, due to memory contention giving diminishing returns beyond 2-threads) threads, which will make me feel less guilty about monopolizing your cycles. :)

[VBCurtis]

Mr Mayer-
I may be willing to donate a couple thread-months of haswell desktop until Luigi gets his machine up and running. If you might be so kind as to PM me instructions for Linux setup (I run ubuntu, if it matters), I can continue a double-check alongside your current run, with the caveat that I'll lose interest after 30-60 days.

[ewmayer]

Originally Posted by VBCurtis

Mr Mayer-
I may be willing to donate a couple thread-months of haswell desktop until Luigi gets his machine up and running. If you might be so kind as to PM me instructions for Linux setup (I run ubuntu, if it matters), I can continue a double-check alongside your current run, with the caveat that I'll lose interest after 30-60 days.

Thanks, Curtis - please, enough with the "Mr.' - you're making me feel my age. :)

First step is to download the current Mlucas release from here and build it - please use the auto-build version at top of the 'Recent News' section. It should end with a binary that is linked to the executable which reflects your platform, e.g. the USE_AVX2 build for Haswell+. Assuming you successfully build & install, run the medium self-tests using 'Mlucas -s m' in a dir created for your work. Fermat-testing needs a few further tweaks, but first let's get through the above.