mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Math > Number Theory Discussion Group

Reply
 
Thread Tools
Old 2017-07-11, 21:53   #23
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2×5,471 Posts
Default

Quote:
Originally Posted by wombatman View Post
From the NVIDIA thread:

ISTR our own David Stanfill (airsquirrels) having successfully downloaded an earlier version of the sw and run it in some very limited way (e.g. to check memory traffic for anything malicious). I just pinged him to see if his GPU farm is up to the author's very demanding standards for required hardware - if David's yuuuge GPU farm won't suffice, then that's a signal to stop wasting anymore time giving this guy the benefit of the doubt.

If OTOH David can run the latest version, we should be able to do some targeted timing tests or start a 'production run' and get iteration times from that, yes?
ewmayer is online now   Reply With Quote
Old 2017-07-11, 22:09   #24
science_man_88
 
science_man_88's Avatar
 
"Forget I exist"
Jul 2009
Dumbassville

26×131 Posts
Default

Quote:
Originally Posted by ewmayer View Post
ISTR our own David Stanfill (airsquirrels) having successfully downloaded an earlier version of the sw and run it in some very limited way (e.g. to check memory traffic for anything malicious). I just pinged him to see if his GPU farm is up to the author's very demanding standards for required hardware - if David's yuuuge GPU farm won't suffice, then that's a signal to stop wasting anymore time giving this guy the benefit of the doubt.

If OTOH David can run the latest version, we should be able to do some targeted timing tests or start a 'production run' and get iteration times from that, yes?
see http://mersenneforum.org/showpost.ph...&postcount=192 new version only needs 1 GPU now.
science_man_88 is offline   Reply With Quote
Old 2017-07-11, 22:16   #25
airsquirrels
 
airsquirrels's Avatar
 
"David"
Jul 2015
Ohio

11×47 Posts
Default

Alright, I’m game. What are the magic requirements for the GPUs needed by this magical program?

I’ll give it a run on the Titan Xp + Dual Titan Blacks three GPU system near me tomorrow. If it is interesting on that 3 GPU system I can try it on a system with 8 K80s (16 GPU cores).

I suspect nothing of interest.
airsquirrels is offline   Reply With Quote
Old 2017-07-12, 01:33   #26
airsquirrels
 
airsquirrels's Avatar
 
"David"
Jul 2015
Ohio

10000001012 Posts
Default

Great news everyone!

Just kidding. This program is very high quality . On a system with 8x Tesla K80s running CUDA 8.0 and the latest drivers the installer for CEMPLLA just crashes. Every time. I disabled all of the accelerators in device manager, at which point the installer runs successfully. Re-enabling the accelerators and launching the application yields a nice error about how the program must not have been installed correctly and you need to reinstall it.

I have a windows computer whose OS install I don't care about running a simple GTX 1060. Perhaps I will give that system a chance tomorrow, but I fear I've wasted enough time playing with this already.

I did dump the disassembly of version 1.5. It still doesn't appear malicious, although it does more work that involves encrypting and decrypting what appears to be its own log and status files. An effort to hide something perhaps?
airsquirrels is offline   Reply With Quote
Old 2017-07-14, 00:13   #27
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2ABE16 Posts
Default

BTW, late last night I uploaded an updated versions of the Mlucas code tarballs (several patches to fix bugs exposed by Alex Yee (a.k.a. Mysticial)'s 10-core-i9 build/tests of the code), plus an updated README page, where I added a new item to the Q&A section at bottom of the page, on how to deal with (apparent) snake-oil pedlars like "Billy Rubin", whose online persona as per the nVidia forums where he peddles his wares seems aptly bilious. LMK if I missed any key "how to gauge the likely veracity of such claims" talking points.
ewmayer is online now   Reply With Quote
Old 2017-07-16, 05:33   #28
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

3·5·7·31 Posts
Default

Quote:
Originally Posted by Prime95 View Post
Dang. I added FFT code to test up to 900M or so. Can't say that I tested it.....
The FFT code may be there, but it errored out right away saying it couldn't be done, so perhaps there's still a check in place for exponents over 596M or whatever size?
Madpoo is offline   Reply With Quote
Old 2017-07-16, 05:53   #29
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

3×5×7×31 Posts
Default

Quote:
Originally Posted by GP2 View Post
This sounds like big news, casually dropped. Will it affect your contributions to the project in a negative way?

I'm a big fan of the cloud, but it allows you to quickly increase and decrease the number of servers to exactly match the needs of the moment, so I imagine there will be less spare computing capacity than before. They may even be planning to lay off sysadmin staff as a cost-cutting measure, since the need for on-site expertise in that area is less after a migration to the cloud.
Oh yeah, losing the physical boxes will be a devastating blow to my own progress, but oh well, I've come to terms with it.

I like going through the work needing triple checks but I have a feeling I can spit out lists of those and other people will happily get on top of those. The other stuff I would normally pick out for myself (strategic double checks) are also good areas where other people already help out quite a bit.

At least right now, I'm mostly caught up on the triple-checking. I've whittled my work queue down to a matter of 1-2 days really, adding in the latest new triple-checks on a "just in time" basis or even poaching triple-checks here and there that won't expire for years but the assignee has given up on it months or even years ago.

As for not needing a (or as many) sysadmin(s), that's probably, generally true in many cases. I know for me personally, dealing with hardware issues is maybe 10% or less of my workload. Actually it's been FAR less over the past year since we haven't invested in any new stuff, it's just been the odd "hard drive died, swap it with a new one".

Many of my functions like dealing with the load balancer and firewall actually translate fairly well to a cloud architecture. The Azure load balancer, for instance, is a VERY basic L4 balancer and isn't even close to what an F5 can do. And setting up firewall rules and ACLs is pretty similar on whatever platform... so instead of Juniper or Cisco I'd be doing it somewhere else. I admit, it is nice not having to worry about the physical aspect of networking... when a firewall or switch dies it's kind of a huge deal, even though I have all of mine setup in redundant pairs.

I still remember doing a Cisco upgrade one late night and of course I managed to muck it up and had to drive to the datacenter to physically reset it. Now I've learned the trick of scheduling an automatic reboot in a few minutes, making my change (as long as it's IOS and not JunOS which has a habit of saving changes to the running config automatically). Then if I mess up, it auto reboots to the old config in a few minutes. If it's fine and I didn't manage to lock myself out, I can save the change then.

Sigh... but all those tricks and skills will soon be lost in time, like tears in the rain. LOL More and more companies just shoving stuff in the cloud and willing to pay more so they don't have to "deal with stuff"...

Maybe I should get a job at a datacenter in their NOC.
Madpoo is offline   Reply With Quote
Old 2017-08-27, 22:07   #30
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

DA416 Posts
Default original topic

From
https://devtalk.nvidia.com/default/t...been-released/

"the amount of time it takes to complete an LLT for a particular Mersenne Prime Candidate may sometimes be less than the time it takes to do the same for a Mersenne Prime Candidate of lesser value. It's not a linear correspondence, as one would expect. In fact, at certain intervals, as one scaled up the binary exponent, and almost completely dependent on the GPU hardware configuration, the total testing time for each candidate would exhibit a 'quantum drop' in value."

This sounds to me rather like some of the CUDALucas issues I've identified involving certain library versions or settings & gpu model combinations: using a larger fft length (which could be triggered by a larger exponent) causes a reduction in total runtime (due to a malfunction of some sort). Invalid residues can be returned very quickly. In some cases I've seen the "orders of magnitude" effect that he references elsewhere, _but it is due to a malfunction_.

Another point to note is shortly before this quote, he references his own limited hardware, implying having done limited testing. It is particularly hard to test such software at the larger exponents, involving longer fft lengths and longer iteration times, when things are working properly, even with rather high-end gpu cards or many fast cpu cores. His software might still effectively be in alpha test.
kriesel is offline   Reply With Quote
Old 2017-08-27, 23:00   #31
CRGreathouse
 
CRGreathouse's Avatar
 
Aug 2006

2×29×101 Posts
Default

Quote:
Originally Posted by kriesel View Post
His software might still effectively be in alpha test.
In the most charitable interpretation, yes.

But the poster hasn't been seen in two months, so I don't know if we'll learn more or get updates.
CRGreathouse is offline   Reply With Quote
Old 2017-08-28, 00:57   #32
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
República de California

2×5,471 Posts
Default

Put me in the "Until he proves otherwise by making his code's interim progress and results transparent to the user, we should assume he is a fraud and not waste our time on any further flummery not related to the foregoing show-us criteria" camp.
ewmayer is online now   Reply With Quote
Old 2018-01-03, 19:19   #33
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

3·5·7·31 Posts
Default

Oh hot diggity... I don't know how I missed this, but he released 1.51 back in September:
CEMPLLA 1.51

Quote:
This new version adds a new utility, called "AddCandy", that enables adding as many Mersenne Prime Candidates as you want to your existing CEMPLLA System Prime Candidate repository. So now you're no longer limited to only those Prime Candidates that CEMPLLA initially comes installed with.
That sounds legit, right?

It sure sounds more and more like this is just some cryptocurrency mining thing... the requirement for lots and lots of GPU, total lack of transparency, etc.

I'm creating a test Windows VM for something else so maybe I'll give this a sandboxed spin just to see what, if anything, I can gather from it, out of curiosity. No GPUs for it to work with, but whatever.
Madpoo is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
CEMPLLA: An alternative to GIMPS ? CEMPLLA Author Data 233 2019-06-28 17:18
380W PSU capable handling 9800GTX+? nuggetprime Hardware 8 2009-01-08 03:33
Is yeafft capable doing IBDWTs nuggetprime Software 5 2008-12-25 12:59
Hello, I'm new to discovering Mersenne primes Vijay Miscellaneous Math 12 2005-04-06 20:11
Credit for discovering Mersenne prime? edorajh Teams 1 2003-11-09 21:52

All times are UTC. The time now is 21:17.

Thu Apr 9 21:17:19 UTC 2020 up 15 days, 18:50, 0 users, load averages: 2.26, 1.94, 1.72

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.