Mlucas v20 available

[ewmayer]

Mlucas v20 has gone live. Use this thread to report bugs, build issues, and for any other related discussion.

Major feature-add: p-1 support
Major ease-of-use add: An automake script makemake.sh

Details at the above link.

Special thanks to Mihai Preda and George Woltman for the total-immersion course in state-of-the-art p-1 stage 2 optimizations, and Paul Underwood for remote access to his Odroid N2 with 4GB RAM, which allowed me to test my stage 2 implementation in an Arm-NEON-SIMD build context. Thanks also to Teal Dulcet for help with the make-Makefile scripting, and Ken Kriesel for trying the automake under WSL.

Please subscribe to this thread if you want to be notified of patch uploads.

[paulunderwood]

On the Arm a53 of my Odroid N2, P-1 takes 8 days for stage 1 and about 10 days for stage 2 using 54.9% of the 4GB of RAM (105Mb numbers).

On the a73 of it a PRP test takes about 6 months.

Ernst, are you gonig to implement "proof certification" soon?

[ET_]

Originally Posted by paulunderwood

On the Arm a53 of my Odroid N2, P-1 takes 8 days for stage 1 and about 10 days for stage 2 using 54.9% of the 4GB of RAM (105Mb numbers).

On the a73 of it a PRP test takes about 6 months.

Ernst, are you gonig to implement "proof certification" soon?

+1

Also, Paul: do your timings refer to a 4-cores multithreaded P-1?

[paulunderwood]

Originally Posted by ET_

+1

Also, Paul: do your timings refer to a 4-cores multithreaded P-1?

No. It is the 2 core chip. The N2 is 2+4 cores: a53+a73. I am running P-1 on the a53 and PRP on the a73.

[tdulcet]

Originally Posted by ewmayer

I see tdulcet's mlucas.sh "do everything" script has not yet been updated, so users of same should edit their copy to change the 2 occurrences of '19.1' on lines 9 and 10 to '20', and the md5 checksum on line 11 to 9b19a56bf9d598a141fd2f655e61f175.

I just pushed the changes to the install/setup script and PrimeNet script for Mlucas v20 to my repository. I wanted to wait until v20 was officially released before doing so. Anyone interested can see the full changes here (scroll down the page). Feedback is welcome.

The install/setup script will also now automatically create a jobs.sh script, as documented on the Mlucas README. It has always dynamically created a Mlucas.sh script, but that only worked when run from a cron job. The new jobs.sh script can be manually run and will not start Mlucas or the PrimeNet script if they are already running.

Our PrimeNet script is now the first and only one to fully supports P-1 factoring with Mlucas, both standalone (the 4 worktype) and when done before a LL/PRP test.

As a side note, one of my systems testing standalone P-1 factoring with Mlucas v20 beta just found its first factor: 105310747.

[ewmayer]

Originally Posted by paulunderwood

Ernst, are you gonig to implement "proof certification" soon?

That's gonna be the major feature add for v20.1. I am hopeful that that will need less time than the p-1 proof-of-principle-work-and-code-it-up did. No fundamental new algorithms or flavors-of-modmul (p-1 needed several new ones of those) needed, just compute a bunch of interim proof-related residues and deposit in files along the way. Teal & Dan's primenet.py is already being enhanced by them to support proofs, so I can focus on the data-processing side of things.

Originally Posted by paulunderwood

No. It is the 2 core chip. The N2 is 2+4 cores: a53+a73. I am running P-1 on the a53 and PRP on the a73.

Doing standalone p-1 on the weak a53 seems like a good use - the a73 can do PRPs in not-completely-hopeless time, but uses very little of your 4GB RAM. Do you have any numbers on how much the a53 p-1 is slowing your a73 PRP testing, due to competition for memory bandwidth?

Originally Posted by tdulcet

I just pushed the changes to the install/setup script and PrimeNet script for Mlucas v20 to my repository. I wanted to wait until v20 was officially released before doing so. Anyone interested can see the full changes here (scroll down the page).

Great - thanks. I'll modify the OP accordingly.

As a side note, one of my systems testing standalone P-1 factoring with Mlucas v20 beta just found its first factor: 105310747.

Congrats! AFAICT yours is the first v20-found factor - 'tis a modest stage 1 factor, but still. All my early beta-test runs were reproducing known stage 1 and 2 factors which I extracted from my gpuowl-on-R7 results files, and past 2 weeks have run a dozen-or-so p-1 jobs on status-unknown expos ~110m on my 2c/4t avx-512 NUC, but none found a factor.

[ewmayer]

Did some unclosed-element cleanup (yes, I am a dinosaur, I still hand-create and edit my html) and spell-checking on the README.html file; also realized that my 'ln -s' example in the Performance-tune section had the source and destination arguments for the link command the wrong way round. Revision uploaded.

[paulunderwood]

Originally Posted by ewmayer

Doing standalone p-1 on the weak a53 seems like a good use - the a73 can do PRPs in not-completely-hopeless time, but uses very little of your 4GB RAM. Do you have any numbers on how much the a53 p-1 is slowing your a73 PRP testing, due to competition for memory bandwidth?

Running (v20) P-1 on the a53 caused the a73 v19 PRP run "clocks" to increase from 00:23:19 to 00:24:49.

[Dylan14]

I've updated the PKGBUILD for Arch Linux to v20 for mlucas. Notable changes:
1. the sysctl-missing.patch is no longer needed.
2. The -lgmp flag is added to the line which begins with gcc -g -o.
3. The gmp package has been added as a dependency.

Note that this still builds it in the way it was done before (check /proc/cpuinfo, run compilation, then check build.log, then link if log is empty) instead of using the handy bash script.

[ewmayer]

v20.1 is out. I suggest users upgrade at their earliest convenience to gain the benefit of various bugfixes, as well as a significantly faster p-1 stage 2.

Please also check your worktodo.ini files for Test|DoubleCheck|PRP assignments mangled by a v20.0 bug. The assignment-repair procedure is detailed in post #2 of the v20.1 thread.