Hi all,

A user testing exponent 15145xxx on a P4 recently got a reproducible round off
error of 0.5.  As we all know that error ruins the LL 
result.  Consequently, I've
changed the FFT crossover point in version 22 from 15,160,000 down
to 15,140,000 on the P4.

In addition, the above uncovered a bug whereby a reproducible roundoff error
greater than 0.48 would cause an infinite loop - backing up to the previous 
save
file in hopes that the roundoff error will eventually become less than 0.48.

To eliminate the problem in the future, version 22 will do round off error 
checking
every iteration for exponents within 2% of the maximum exponent that can be 
tested
by the current FFT length.  The current version only does roundoff error 
checking
every 128 iterations.  If a reproducible roundoff error is greater than 
0.45, then
we restore the previous save file and shift the data one bit - this should 
change
the FFT data so that a more acceptable roundoff is generated.

What should you do if you are running version 21 on a P4?  You can do one
of the following:
1)  Nothing.  You are unlikely to run into trouble.  Just keep an eye on your
results.txt looking for an unending string of roundoff errors on the same 
iteration.
2)  Turn on Advanced/Round Off Checking.  This will error check every iteration
instead of every 128 iterations.  Again keep an eye on results.txt for unending
error messages.  Roundoff checking on the P4 is very fast.
3)  Upgrade to version 22 (only the Windows version has the above fix).  It 
isn't
fully tested but should be safe.  The only dangerous change I've made is 
compiling
with maximum optimization to speed up the GCD code.

Sorry for the trouble,
George

_________________________________________________________________________
Unsubscribe & list info -- http://www.ndatech.com/mersenne/signup.htm
Mersenne Prime FAQ      -- http://www.tasam.com/~lrwiman/FAQ-mers

Reply via email to