At 10:56 AM 5/16/2001 -0000, Brian Beesley wrote:
>On 16 May 2001, at 0:24, Ken Kriesel wrote:
>
>> For the time being I would like to continue focussing our QA efforts on the
>> general case, rather than P4's specific limits, since
>
>Agreed. I was talking about the "short" test suite, where the runs
>stop after 400 iterations. We can generate the required data for a
>large number of exponents without diverting significant effort. In
>fact it's much easier to do this than it was a couple of years ago,
>as we have at least two reasonably efficient programs other than
>Prime95 which can be used to generate & cross-check the data.
>>
>> If I recall correctly, this does not require "ridiculously" small
exponents,
>> since the convolution error falls off quickly away from our usual upper
>> limit on exponent for a runlength.
>
>Quite correct. IMO it is "ridiculous" to run a whole LL test using a
>FFT run length of 2R when a run length of R would suffice; but
>running at least some 400-iteration short tests with a "half size"
>exponent during a self-test does make sense in two ways: it saves
>having to generate quite as much short run data (since we can recycle
>some of the smaller exponents into larger self-test run lengths), and
>- providing there is a temporary reduction in the rounding error
>limit whilst these short runs are in progress - we would be making
>the self-test more stringent.
Yes. Even the next-longer runlength provides a very large drop in
the expected convolution error for those exponents near the limit
of a runlength.(orders of magnitude.) Two to one on runlength is not needed.
The ratio 5:4 gives substantial reductions.
>There is no point in optimizing the self-test for speed, since we are
>going to run the self-test for a fixed period of time.
There is some point in having the self-test resemble closely, actual running
conditions, to be as valid a test as possible, with respect to everything from
bit patterns to power loadings.
Ken
_________________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
