Improved params files, 2022 edition

[VBCurtis]

Attached are new params files for C95 through C130.

I'm still confused about my test results for C135-C145, so I continue to test.

Params for C150+ jobs are scattered around this subforum- once I get good drafts posted up to C145, I'll collect the best guesses we have so far on C150+ into a new set of files in this thread.

I welcome A/B testing of factory settings and these settings- if we show clear speed gains on a variety of hardware, then we can submit these to the CADO group for inclusion as future "factory" settings. A summary of wall-clock time for factory vs new params file is all we need on these small files.

[VBCurtis]

Wall-clock timing comparison on a Ryzen 5950, CADO version current git as of May '22:
RSA-100 Factory 343 sec Improved 180 sec #Factory poly search is deg 4
RSA-110 Factory 536 sec Improved 555 sec #hrmmm, improved?
RSA-120 Factory 2016 sec Improved 1678 sec

[swellman]

A/B testing on an idle i7-3770S CPU @ 3.10GHz (Ivy Bridge), 32 Gb RAM, Windows 10 using Ubuntu 18.04 LTS, CADO installed late Aug 2022, using random composites found in factordb:


c90
571182163113957287170179806750737837113436761259235360848031582728863975649824460077229603

Factory - 2683 seconds, Improved - 1777 seconds

c92
44577254493130764909193544753471835604872768288806087823258523406159832841077062167087180397

Factory - 3275 seconds, Improved - 2173 seconds

c95
61678211188677863570157133010819984802060134393238837081832540114742617862791150282895502269369

Factory - 4152 seconds, Improved - 2930 seconds

c100
3026995565521796176841046970188026667837386421497914848140699810836615249355620806152722969475594003

Factory - 8922 seconds, Improved - 6234 seconds

c105
349146070982960202561848166250693288094792170771592377901684145676723080820817407502131216515530304977117

Factory - 14827 seconds, Improved - 10495 seconds

c110
54559505794893210575630412396836182993659417080441518746765185717148940515062581106037049619993173736599367017

Factory - 24974 seconds, Improved 22697 seconds

c115
2693818177394916277639013170129551006330371999162847185623364054776531064764732471565819889978176436301757113870687

Factory - 49954 seconds, Improved - 39247 seconds

c120
217221355657878614893196197864671541103658421873582813009356933764214657538563824989226987509276923860046855162176551141

Factory - 84009 seconds, Improved - 70669 seconds

c125
69921965218695300207302836905717351598843625309602892595999904890558896412757214218189594809196012224191740893445726272198253

Factory - 177016 seconds, Improved - 138895 seconds

c130
4274423118885919346909105300287512500348793228681071392161725722181403752611150880697482969247877547695528710043215870640174290209

Factory - 362497 seconds, Improved - 257333 seconds

[VBCurtis]

Here are "draft" files for C135 through C150. Timings using these files are within 5-7% of the trendline established by C100-C130 files in post #1, and I'll keep playing with settings to try to get that last few percentage points.

This .zip file will contain the most up-to-date settings I have; if I update a file I'll change the date in the .zip name.

EDIT: Jan '23 I think I've hit the trendline for C135-C155, and a variety of small changes yielded no further speed improvements.

[chalsall]

Originally Posted by VBCurtis

Here are "draft" files for C135 through C150.

Please forgive me for this... But...

Are we actually passing parameters around between humans?

Arguably doesn't scale.

[VBCurtis]

I don't know what you're asking, Chris.

If you mean you don't know why they're distributed here rather than with CADO, I'm working on that.

[chalsall]

Originally Posted by VBCurtis

I don't know what you're asking, Chris. If you mean you don't know why they're distributed here rather than with CADO, I'm working on that.

Sorry... I was channeling my inner Inspector Clouseau... 9^)

I guess fundamentally I'm wondering how far along the automation of Work Units for all the various different projects is/are.

I've been a bit spoilt by PrimeNet and its API. Remarkably advanced, even by today's standards (often better).

I joined GIMPS after Primenet was online and the clients were doing IPC. Before that, I understand George used to distribute WUs by way of emails.

A similar thing used to be done with TF'ing when the GPU code paths first manifested.

Never send a human to do a machine's job...

[VBCurtis]

Post #4 contains updated params for C135 through C155. There may still be a bit more to find at C150, as C155 is below the trendline but C150 is not. I'd rather explore C160+ this year.

At C160, 3 large primes start to sieve faster but with a much larger matrix- I think at that size whether to use 3LP depends on whether one is using msieve for the matrix (use 3LP) or CADO (use 2LP to save matrix time). That tradeoff depends also on machine architecture- some are relatively faster at matrices than others. Once I do more testing, I plan to post fastest 2LP sieve params as well as fastest 3LP params, so individual testers can try both and choose.

[VBCurtis]

Attached is a C160 params file using 3LP, intended for those who use msieve for postprocessing. These params sieve about 25% faster than my best 2LP, but more relations are needed to build a reasonable matrix.
Folks like Ed who use a GPU for matrix-solving and just want to get a matrix as early as possible will likely find 225-240M relations are enough. Folks running msieve on CPU for LA may wish to experiment with rels_wanted: I'm using 255M based testing of just two jobs. C158 should get a matrix around 6M, while C161-162 will likely be around 7M size with msieve TD of 100-110.

Those using CADO start-to-finish might want to adopt the poly select from this file, but use 2LP sieve settings to get a matrix that solves in less than 1/3 the time it takes to sieve. Using the attached 3LP sieve settings with CADO for matrix should target 270M relations (or more- I tried 260M for a C158 and decided more sieving would help). Msieve matrix-solving is nearly twice as fast as CADO for the same relations set- at this size, that's a meaningful time savings.

I have not yet done a thorough test-sieve with a C165 to make the next-size-up params file, but a reasonable guess would be to add 25-30% to each lim, add 1 to mfb1, add 0.03 to lambda1, add 5-10% to rels_wanted, and add the same percentage to q-min that you added to lim1 (that is, make qmin a bit under 20% of lim1). When I start testing C165s, I'll try these settings against whatever we've developed in the c165-175 thread.