At 07:07 PM 6/22/2008 +0200, you wrote:
>Ken Kriesel wrote :
>
>> For a hypothetical pool of machines of constant reliability,
>> the error rate per trillion operations (or should that be per
>> hour of 100% cpu utilization?) would be about constant.
>>  The number of operations goes up slightly faster than the
>> square of the exponent, so twice the exponent means more than
>> 4 times the error rate per exponent.  (More iterations,
>> performed in more pieces, of slightly lower precision)
>
>George Woltman increased the reliability of the software to decrease the
>error rate (more checking during the LL tests...)
>
>> It's also possible to have an error in factoring attempts
>> yield a false positive for a factor that fails verification.
>
>Factors are verified when submitted, according to a post by George
>Woltman on the Mersenne forum. What could happen are missed factors due
>to hardware errors.

My recollection is I've seen the false positive factor case occur 
(in instances not using primenet).  Of course the other factoring 
error occurs silently (missing a factor that if computation occurred 
correctly would have been found).

>> If I recall correctly, verification of a Mersenne prime is 
>> run on different software on a different computer
>> architecture to maximize the chance of a software bug or
>> computer hardware design error from yielding a false
>> positive.  During the QA effort, comparisons of interim LL 
>> test residues were made among differing software run on
>> differing computer architectures.
>
>This is only done when a LL test finds an exponent yielding a prime.
>Otherwise double-checks are accepted even if done by the same machine
>with the same software version if the different shift gives the same 16
>byte residue.
>
>Jacob

Yes, verification of a Mersenne prime is a very special case double-check,
so it is done differently than the double check of a negative result.


Ken
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to