On Tuesday 07 November 2006 12:18, david eddy wrote: > > Why "rare" instead of "optimal"? Well, I don't want to be the one to > > tell someone > > he missed a prime because his run had a not-unexpected round off error. > > OTOH he (or she!) was more likely to miss finding a prime because excessive > time was wasted using the larger FFT.
Umm. If you get the wrong residue you're wasting all your time. Also if you're cutting it too fine, you will start getting a significant rate of roundoff errors detected, which may be recovered but "waste" time re-computing from the last savefile. Implementing full roundoff error checking costs about the same in terms of time "wastage" as going to the next run length. In any case a detected roundoff error of 0.5 cannot be reliably recovered from as it's impossible to guess accurately whether the correct integer is the "floor" or the "ceiling". > > It amounts to paying unnecessarily high insurance. > How about informing testers of exponents at the bottom of one of > your ranges what the risk of using the smaller FFT is, and allowing > them to take it if they want to? It may be appear to be "optimal" to guess a residue and report that without bothering to run any of the test! After all there's a 1 in 2^64 chance you will get it right... The server wouldn't react well to this if everyone tried it, though. > > BTW I replaced Prime95v24.14 with v23.18 mid run, achieving a 14% > speed up on my PII. Any danger I have sabotaged the test? Should be OK, the most likely problem with a reverse upgrade is compatibility of the savefile. There is however a problem with some exponents when switching between computers - even with the same software version. If you switch from a non-SSE2-capable processor type to a SSE2-capable processor type, and the exponent is between the SSE2 and non-SSE2 run length limits, you will almost certainly end up with excess roundoff errors, and very likely with a wrong residue - even if all the roundoff errors were apparently recovered. Regards Brian Beesley _______________________________________________ Prime mailing list [email protected] http://hogranch.com/mailman/listinfo/prime
