Team Ninja Bulletin Board  

Go Back   Team Ninja Bulletin Board > DC Vault > Project Removal

Reply
 
Thread Tools Display Modes
  #1  
Old 18th December 2015, 10:45 PM
Rusty Rusty is offline
Owner
 
Join Date: Jan 2004
Location: Tasmania, Australia
Posts: 11,388
GIMPS(cheating stats)

Really? Has it got that tough out there in stats land that people have to manipulate results to get extra points..

Very sad.. Very sad indeed.

Well in my book this still falls well within the rules however I am not comfortable(like the moneybee discussion) with idiots tampering with results to gain a few more points and more to the point the project is doing nothing to improve the outdated system.



Hit me up!

Thoughts?
__________________
RUSTY


Team Ninja Forever : Once a Ninja, always a Ninja - Team Ninja

"I'm a SAS NINJA"

Drafted to the SAS
Dump of the day in Folding@Home   Dump of the week in Folding@Home   Dump of the month in Folding@Home  
Reply With Quote
  #2  
Old 23rd December 2015, 04:39 PM
Coleslaw Coleslaw is offline
[H]ard|OCP
 
Join Date: Oct 2013
Location: IL
Posts: 594
I have read where people are curious about this thread. Here is a forum link with some insight on the issue. http://www.mersenneforum.org/showthr...=20438&page=11 We had noticed a large stats change that was brought to our attention by a forum post by a member from OcUK. This is something that has happened more than once. The stats had been adjusted and work appeared to get double checked. However, it appears to be a very well known flaw to their scoring system in that there is no real idea how much cheating actually takes place as long as it is done in small amounts to not tip someone off. I'm not well versed in the inner workings of this project as I was never a heavy donor or even that interested in mathematics. However, the issue of the cheating potential is still there and Rusty is essentially asking for others to dig in a little bit and give their honest opinions. Something like this is needed to be discussed as it could very well affect whether teams want to support it in a "challenge" or at all. So, please take a moment to think it over. Think past your teams over all past and current contributions to the project and consider whether not this behavior whether at this project or any other is tolerable. Then discuss it with your teams and finally report back here and discuss for others to consider. The more teams/users that chime in the better. It is not fair to us and others for your team members to drag their feet now and then complain later about a decision being made. If you truly have an opinion, let it be heard.
__________________
[link=http://join.worldcommunitygrid.org?recruiterId=338542&tea mId=BP5XNJBR9N1][/link]
Reply With Quote
  #3  
Old 25th December 2015, 01:20 AM
iconized iconized is offline
Ninja Jotohei
Dutch Power Cows
 
Join Date: Jun 2011
Location: Netherlands
Posts: 68
Tried to start a little discussion/information exchange about it on the DPC forum, but no one has responded yet.
I will look into it, but for now I have no idea how the cheaters operate. Could be something like claiming false factors.
There is this expression like: "Throwing the baby away with the bathwater"
I like GIMPS and yes they (the project management) should do whatever possible to avoid or punish cheating.
Meaning we (Vaulters) should also voice our opinions on this on the Mersenne forum: http://mersenneforum.org/
Reply With Quote
  #4  
Old 25th December 2015, 08:58 PM
Coleslaw Coleslaw is offline
[H]ard|OCP
 
Join Date: Oct 2013
Location: IL
Posts: 594
I agree people should post in their forums to get a good feel for how important it is to everyone. However, there are teams that don't participate at GIMPS and probably wont care to take the issue to them in the first place. So, please post here as well as this is also to address how we should handle things at the Vault. The Vault wont dictate how other projects operate. We can only bring it to their attention.
Reply With Quote
  #5  
Old 28th December 2015, 10:41 PM
VictordeHolland VictordeHolland is offline
Neophyte
DPC member
 
Join Date: Jul 2011
Location: Holland
Posts: 15
You can't upload false factors to GIMPS, the server checks for that.
If it is a factor of an exponent, than it should be (numbertofactor = 0 modulo factor). It is a very inexpensive operation for a CPU/server to check, what does take the server some time is to check the factor reported is new (read: not in the database already) and to check it's primality (a PRPtest or Primo will take care of that).

The obvious cheated results have been deleted and/or verified by others.
There is no system in place to catch every bad result, but there is no need for that.

There are 4 different work types on GIMPS: LL/DC, Elliptic Curve Method (ECM), Trial Factoring (TF), P-1 Factoring (P-1)

LL/DC
Every first time LL test is verified by a other computer/user (called a Double Check or DC test). Using different shift values (to prevent software bugs from causing identical results) the DC test will catch all the faulty/bad LL tests. If the residues don't match a TC (Triple Check) will determine the correct results). About 3-4% of the tests are proven erroneous/faulty this way. I would say probably 99% of those bad tests are caused by faulty hardware (overclocked CPUs/ bad memory). The rest by faulty software/bugs, upload errors and maybe the occasional cheated results (why you would want to cheat a LL test that could possibly be a record breaking million digits earning you a few $1000 is beyond me!)

For : ECM, TF and P-1 there is no such system in place. But you will draw suspicion if you don't regularly find factors by TF. You should find about 1 factor in 70-80 tests for TF before first time LL and 1 in 100 tests for TF before DC (because P-1 found some factors). If you do TF on very small candidates that have had some ECM or lots of P-1 (like me) the chance drops dramatically.
It just doesn't make sense to double check all the TF work for a couple of reasons:
1. It would halve the TF throughput of the project!
2. The cost of a missed factor is not very high: You have to do a LL test and a DC test on that exponent. A factor or matched LL/DC tests is the same for the project goal. It determines the number is not prime!
3. The factors found to TF test preformed matches the theoretic probability of factors. The missed factors don't add up to a very significant amount.

ECM is a bit different, you can easily find additional factors for exponents that have already a known factor. That is the reason some people have thousands of factors for ECM, while others who do ECM on exponent without a know factors (which you don't know the size of the factor yet) have 0 or only a few factors found. The amount of ECM work needed to have a significant chance of finding a 40 digit factor (if the exponent has such a factor) is tiny compared to that of a 65 digit (3 orders of magnitude).

People can do ECM with GMP-ECM for stage 2 (it is more efficient than Prime95 for small exponents (<10,000 digits) and email George (the project admin) the results/logs who will import them in the database. That explains the occasional 10,000+ GHzdays added to my and various other accounts). I can send you my GMP-ECM logs, if you want to check my latest batch of ECM on M1277 .
In the end it is an apples to oranges comparison between TF (very efficient on GPUs) to LL (CPUs with AVX) to ECM (most efficient with stage 1 done by Prime95 and stage 2 by GMP-ECM).
The total amount of ECM work on the project is not that much compared to TF and LL. The Top 25 teams have only a few percent of their work done as ECM, many have 70-80% as TF.

You can't compare BOINC credits between projects either, this is not that different in my opinion.

If you have any further questions about GIMPS, feel free to ask them. I rarely check this forum, so it can be a couple of days before I read them.

Last edited by VictordeHolland; 28th December 2015 at 10:54 PM. Reason: clearing things up
Reply With Quote
  #6  
Old 29th December 2015, 01:08 PM
Coleslaw Coleslaw is offline
[H]ard|OCP
 
Join Date: Oct 2013
Location: IL
Posts: 594
I don't think anyone here was comparing the points to BOINC or like BOINC... The issue was brought to our attention at the time because a few members had trillions of GHz days in the matter of a day or two. And then reading in that forum suggested that even cheating on a low scale was tolerable. So, that brought up the discussion on how the DC-Vault should consider such projects. With BOINC, many have a double checker option. No, not all, but most. Most projects investigate the cheating and take action against them and try to fix the issue or make changes to prevent it from being possible. GIMPS adjusted those users stats of course. But as far as I know, they have done nothing to prevent the problem on any scale from happening again.

Edit: And thank you for the write up on the project as there are many people out there that don't understand it as is.

Last edited by Coleslaw; 29th December 2015 at 01:43 PM.
Reply With Quote
Reply

Bookmarks

Thread Tools
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is On
Forum Jump


All times are GMT. The time now is 02:26 AM.


Powered by vBulletin® Version 3.7.3
Copyright ©2000 - 2018, Jelsoft Enterprises Ltd.