On Tuesday 17 January 2006 21:30, Tim Sloane wrote:
> >
> > Probably that your machine gets overheated. Check the temperature of
> > your processor

Reproducible errors, very unlikely to be a hardware problem.
>
> Actually, I believe this type of error comes from the Fast Fourier
> Transform (FFT) length not being big enough.  In Prime95, it occasionally
> checks the round-off error from doing the FFT required for the
> multiplication of large numbers during the Lucas-Lehmer test.  If it's
> rounding off more than 0.4from the resultant numbers, it switches to a
> longer FFT length
> which is more
> precise, but takes longer to test.

About right. The problem is that the round-off error is distributed 
approximately normally; very occasionally there is a rogue set of data - 
resulting from correct operation of the correct algorithm - which results in 
an unusually large round-off error. 

Towards the top of the range of exponents for any particular FFT run length 
there is always a small risk of an excess round-off error. If we're lucky 
then we catch it & generate the output seen. If we're unlucky then we might 
miss it. 
>
> So the error is a most likely due the algorithm, not your hardware.  If it
> was a hardware problem, it is likely that the error would not be
> reproducible, which it was.

Not the algorithm's fault - but the way in which we implement it. We could be 
ultracautious and use a larger FFT run length, or even just turn on round-off 
error checking every iteration, but there would _still_ be a finite chance 
something would get missed, and the whole project would run significantly 
slower.

Double-checking with a different offset makes the data input to the FFT 
different at every iteration, so a double-checked residue match is reliable. 
A missed excess round-off error on either run - should be very rare - will 
cause a mismatched residue and result in an additional run being necessary.

Regards
Brian Beesley

p.s. isn't this in the FAQ?
_______________________________________________
Prime mailing list
[email protected]
http://hogranch.com/mailman/listinfo/prime

Reply via email to