Mersenne: Round off error = 0.5
Hi all, I'm running Prime95 on a PII-400 for 6 days (no overclocking) at exponent 9409271. It's produced several outputs concerning ROUND OFF ERRORS - the last one is ROUND OFF [0.5] 0.4 What to do now? Restart from iteration 1? Regards Dieter Schmitt
Re: Mersenne: Round off error = 0.5
Hi, At 01:04 AM 1/25/00 +0100, Dieter Schmitt wrote: I'm running Prime95 on a PII-400 for 6 days (no overclocking) at exponent 9409271. It's produced several outputs concerning ROUND OFF ERRORS - the last one is ROUND OFF [0.5] 0.4 What to do now? Restart from iteration 1? The first thing to do is try and figure out if your CPU is overheating or you have some flaky memory chips. You almost certainly have a hardware problem of some type. Prime95 will go back to the last good save file after the error. So there is a chance your result is OK. The question is "did a hardware error go undetected?" I have little hard data on this, but I'd guess several errors of this type mean you have less than a 50-50 chance of producing a good result. Regards, George _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: Mersenne: ROUND OFF error
Hi all, At 08:13 PM 7/20/99 -0700, Daniel Swanson wrote: Iteration: 6128000/7812379, ERROR: ROUND OFF (0.4026489258) 0.40 [Tue Jul 20 16:35:33 1999] Disregard last error. Result is reproducible and thus not a hardware problem. I consulted the readme file, but it wasn't very helpful. What is a "ROUND OFF" error? While it's nice that it's reproducible, will it invalidate the result of this LL test? Does the fact that this exponent (7812379) is so close to the FFT size breakpoint (782) make this type of error more likely? Yes, the fact that you are testing an exponent near the FFT size breakpoint makes this error much more likely. It will not invalidate the result of your LL test. Note that in version 19, I've decided to be a bit more conservative with the FFT breakpoints. Your exponent will be double-checked with a larger FFT size (if the double-checker uses v19). Regards, George P.S. Double-checking has not revealed any problems near the upper ranges of any FFT sizes. The more conservative FFT breakpoints will reduce the chance of an incorrect result due to an excessive convolution error. _ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Mersenne: ROUND OFF error
I got the following set of messages today: [Tue Jul 20 16:15:19 1999]Iteration: 6128000/7812379, ERROR: ROUND OFF (0.4026489258) 0.40Possible hardware failure, consult the readme file.Continuing from last save file.[Tue Jul 20 16:35:33 1999]Disregard last error. Result is reproducible and thus not a hardware problem. I consulted the readme file, but it wasn't very helpful. What is a ROUND OFF error? While it's nice that it's reproducible, will it invalidate the result of this LL test? Does the fact that this exponent (7812379) is so close to the FFT size breakpoint (782) make this type of error more likely? Thanks, Dan Swanson
Re: Mersenne: ROUND OFF error
At 08:13 PM 1999/07/20 -0700, you ("Daniel Swanson" [EMAIL PROTECTED]>)wrote:
I got the following set of messages today: [Tue Jul 20 16:15:19 1999]
Iteration: 6128000/7812379, ERROR: ROUND OFF (0.4026489258) > 0.40
Possible hardware failure, consult the readme file.
Continuing from last save file.
[Tue Jul 20 16:35:33 1999]
Disregard last error. Result is reproducible and thus not a hardware problem.
I consulted the readme file, but it wasn't very helpful. What is a "ROUND OFF" error? While it's nice that it's reproducible, will it invalidate the result of this LL test? Does the fact that this exponent (7812379) is so close to the FFT size breakpoint (782) make this type of error more likely? Thanks, Dan Swanson
The calculations are done in the fft using an irrational base, then converted
back to integer form by rounding. The amount of rounding is monitored. Well
away from the size breakpoints, the roundoff may only be thousandths, or less.
The FFT size breakpoints are determined by estimates of how high an exponent
can safely (without roundoff error) be run on the shorter faster size.
When roundoff exceeds 0.4, it is possible that a memory error or cpu error
caused it. It's also possible that the roundoff was actually greater than 0.5
and rounding to an incorrect integer has occurred. Exponents near fft size
breakpoints, or any that generate both the roundoff error and the reproducible
message are therefore somewhat more likely to fail a double check, and so are
candidates for earlier double checking.
Nonreproducible roundoff errors mean hardware may be unreliable;
reproducible roundoff errors mean fft size breakpoints ( LLtests) may be
unreliable.
Ken
_ Unsubscribe list info -- http://www.scruz.net/~luke/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
