I might have delivered some bad results

[iconized]

I have done some GPU TF for GIMPS since half september this year and for reasons I already described elsewhere the results might not be so trustworthy:
http://mersenneforum.org/showthread.php?t=16227

Sorry for the inconvenience,

iconized, DPC.

[Dubslow]

Check the other thread.

[Dubslow]

Call for help: There are roughly 2000 exponents, and that will take me the better part of a month for me. Anybody willing to take some of these? (iconized, if you've since fixed your memory, could you redo some of them?)

[iconized]

Yeah I already warned you in the PM it is quite a lot unfortunately. There could be other solutions though, like removing the results from the database so it will be automatically offered again as new TF candidates. I should lose the GHzdays also then.

Problem here is I can't be sure when the memory module started to malfunction. From halfway September I ran RNA World (3 cores) together with Mfaktc (1 core). RNA World uses huge amounts of memory so the danger zones of 2-4 GB and 6-8 GB must have been used. I would expect that application to crash on me when the memory malfunctions badly, but I am not sure.

In October it was PSP-PRP (3 cores) and Mfaktc (1 core), that fits easily in 2 GB together with Win7. In November I have only done PSP-PRP and that also fits in 2 GB. In DPC there is a group of people called the Vault Hunters who jump projects almost monthly and I am part of that.

On 5 November I noticed some odd things while extracting a 7 GB file downloaded from d2d, same time I had removed the Windows swap file (with 16 GB not really needed). It repeatedly gave me different CRC errors while unpacking. Downloaded again same story. Now with the faulty module removed I have concluded that both files had download errors and unpacking errors. There were files in that zip like 1 GB big so the bad memory was used.

So what's wisdom? I am willing to help with the recheck of course, but perhaps only the found factors need rechecking? False factors can be damaging to the project. I can do that myself but I need a judgement call from the project leaders.

[retina]

Originally Posted by iconized

...

RNA World uses huge amounts of memory so the danger zones of 2-4 GB and 6-8 GB must have been used.

...

In October it was PSP-PRP (3 cores) and Mfaktc (1 core), that fits easily in 2 GB together with Win7. In November I have only done PSP-PRP and that also fits in 2 GB.

...

There were files in that zip like 1 GB big so the bad memory was used.

With the paging mechanism used in all protected OSes there is no guarantee what physical memory addresses are being used. It can be that even a seemingly contiguous 8kB memory block (contiguous in the virtual memory view) will be allocated as two independent 4kB pages in different physical memory modules at non-contiguous physical addresses. So for your situation ALL programs you run while having bad memory are possibly subject to using the faulty modules. Don't trust any work from that machine no matter how large or small its memory footprint is.

[markr]

Originally Posted by iconized

So what's wisdom? I am willing to help with the recheck of course, but perhaps only the found factors need rechecking? False factors can be damaging to the project. I can do that myself but I need a judgement call from the project leaders.

No need to check found factors - the primenet server checks all factors submitted to it. Unless you want to be sure all factors were found in the "bit range" you tested.

[Dubslow]

I was planning on going through and verifying that PrimeNet did accept them. I should have everything done and ready for work later today.

iconized, the best way to get the data off of PrimeNet is to PM George Woltman (username Prime95) with your results file and the description of your problems.

[kladner]

...there's around 2000 TF that needs to be done, and I'd like some help

Dubslow-

Regarding the above quote from here:
http://www.mersenneforum.org/showthr...=16104&page=11

Are you saying you want GPU assistance? I can throw some mfaktc time at it if that's what you're looking for.

[Dubslow]

Indeed, that's what I meant. How many do you want? Expect a PM within 24 hours.

[kladner]

Originally Posted by Dubslow

Indeed, that's what I meant. How many do you want? Expect a PM within 24 hours.

Why don't you start with 90-100, depending on what you can fit into a PM. I'll shuffle things to get one worker on them immediately, and add more once the current assignments clear out.

[cheesehead]

Originally Posted by iconized

Yeah I already warned you in the PM it is quite a lot unfortunately. There could be other solutions though, like removing the results from the database so it will be automatically offered again as new TF candidates. I should lose the GHzdays also then.

Don't worry about it!

PrimeNet has had various safeguards built into it for a long time. (After all, what about the thousands of erroneous results reported over the years by other folks who didn't realize their results were incorrect???) GIMPS won't be harmed if you don't personally go back to correct your past reports -- there are already automatic mechanisms for others to do that, and you don't have any personal obligation to re-do what you've submitted in the past.

So what's wisdom?

Breathe deep, then relax. Carry on fixing the problems on your end, but don't worry about PrimeNet or GIMPS -- they can protect themselves from erroneous results.