On 6 Mar 00, at 23:09, Ken Kriesel wrote:
>
> Lots of results (verified by 2 independent runs on separate processor
> generations in V19 prime95) were available at iteration counts of 100 and
> 400 if I recall correctly, partly for validating future programs, at
> ftp://lettuce.edsc.ulst.ac.uk/gimps/PrimeQA/QADATA.TXT thanks to Brian
> Beesley for generating the list of exponents to run on, running most or
> all of one pass, and hosting the data. I see that file's no longer there.
Actually the file contains the low 64 bits of the residual after 400
and 1000 iterations for smaller exponents (up to 4 million), 400
iterations only for medium size exponents (4 million to 20 million)
and 100 and 400 iterations for larger exponents (20 million to 79
million).
The results were created on an Alpha 21164 using a version of Richard
Crandall's "original" LL testing program lucdwt, modified to write
the interim residual & stop instead of completing the test. It's very
slow compared with the current production software, but much simpler!
I verified the results obtained against both Prime95 v19 on Intel PII
(during beta testing, using both the "classic" and "PPro" tunings of
the FFT code) and Mlucas v2.7z on the Alpha system.
I've just put the file back up. Sorry, it got removed accidentally
when I upgraded the server hardware about three months ago.
[ Ernst Mayer comments ]
> >So, it doesn't look like testing exponents above ~40M is going to
> >be practicable any time soon, where I mean doable in a year or less. But
> >since the vast majority of GIMPS first-time LL tests won't even be
> >approaching 20M for some years yet, having one or two double-checked
> >exponents in each subrange below 39M would seem sufficient for the next
> >couple of years.
I agree. Of course, anyone who thinks it's _fun_ to tie up a system
for several years running a LL test on a larger exponent is quite
welcome to do so, so far as I'm concerned!
> I'd like to see them get cpu credit, but I am not in a position to
> guarantee it.
George seems to add the CPU credit from QA tests to his records. So
far as PrimeNet is concerned, we (deliberately) don't use PrimeNet to
communicate results; in any case PrimeNet doesn't recognize that we
own the QA assignments (some of them are actually triple-checks & the
rest are outside currently active ranges), which is why they "don't
count". This doesn't bother me, but it should be easy enough to fix.
> Most of the effort would fall on someone else, and I'd rather see George
> and Scott doing other things than the bookkeeping of apportioning credit
> by iteration count and exponent size and checks of usable save files. To
> keep the minimum contribution sizable, I ask the volunteers to commit to
> at least a half-PII-400-year; large contributions are more likely to
> justify crediting the work or a partial large exponent.
>
I'd strongly reccomend anyone running PrimeNet 10 million digit
assignments to place the following line in their prime.ini file:
InterimFiles=1000000
This will cause the interim residual to be written to results.txt
every 1000000 iterations. At the same time, an extra save file will
be written. The idea is that, when double-checking is scheduled on
this exponent, any error can be found without neccessarily having to
complete the whole run to find it, thus saving time. Also, in the
event that a prime is found, with a set of interim save files we will
be able to verify the result much more quickly by running 1000000
iterations on thirty odd systems in parallel.
If you're worried about disk space, delete the extra save files; the
interim residual has considerable value in itself.
Also could anyone dropping out of a large exponent run (including
PrimeNet 10 million digit range assignments) please send in a copy of
their last savefile, so that work completed isn't lost. My server has
lots of space available for this task.
Regards
Brian Beesley
_________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
