On Monday 16 June 2003 20:16, George Woltman wrote: > > I'm also adding code to 23.5 to check EVERY iteration for an impossible > result such as -2, -1, 0, 1, 2. This test will be very, very quick.
Sounds sensible to me ... but, does it not make sense to run this test during those iterations when testing for excess roundoff error occurs, i.e. normally only 1 in 128 iterations? The point here is that, once the residual gets into a loop like this, it won't get out again. > > FYI, six times a result of 0000000000000002 has been reported to the > server. So, somehow or another it is possible for a hardware error to > zero the FFT data without triggering an ILLEGAL SUMOUT. > Any instances of FFFFFFFFFFFFFFFE? Should be about as common, if this is a hardware related problem. There are lots of reasons why memory corruption may occur but, in most cases, it is hard to see how a whole block of data should be filled with zero (or one) bits without corrupting the program code in such a way that the program would crash or have the operating system crash from under it. I think the most likely scenario would be that a pointer to the FFT work space could be corrupted & point to virtual memory which has "zero on demand" attribute. This might be detectable by memory leak even on systems without proper memory protection (Win 9x) but could be fixed easily enough by keeping _two_ pointers to critical work space (the values don't change that often...) & comparing them occasionally. As to why the pointer might get corrupted, most likely we're looking at stack overflow or some other program behaving badly rather than a bug internal to Prime95. It would be interesting (though probably impossible) to check which OS the "residue 2" runs were run on. If my logic is right then I would suspect that they were all run on Win 9x/ME rather than NT, W2K, XP or any of the varieties of linux, where proper memory protection should give much better protection against this sort of problem. Regards Brian Beesley _________________________________________________________________________ Unsubscribe & list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers