Since the release of v17 supports double checking of results
obtained by other Prime95 machines, is it possible to set up the
server so that it assigns each number out to two clients? One could
be picked, arbitrarily, to be doing an old-style LL test, and
another could use the new double check code. It seems that this
might be a good way to address the issues raised when some numbers
are predicted to take a huge amount of time to finish.
This would have to be done intelligently, otherwise we'd cut
the number of exponents tested once by a significant amount (under
the assumption that most exponents will be checked once and returned
in a timely manner). It should probably be restricted to numbers
which will take longer than 60, 90, 120 or however many days we deem
to be "a long time" on the machine they are first assigned to. This
limit might even be dynamic, depending on how low the exponent is,
assuming that we want to fill in the gaps in the lower exponents
quickly.
There are a few drawbacks. First of all, what happens if
one of these exponents turns out to be the next prime. If computer
A checks out a number, works on it for 3 months, and then computer B
gets assigned the same number to double check and completes the test
before computer A, is it fair to "take away" computer A's discovery?
On the plus side, there isn't an issue of "stealing" credit for
composite numbers, assuming that x CPU years of LL tests == x CPU
years of double checking exponents --- all that changes is the order
of checking things in (the double checking actually finishes before
the original LL test).
Also, this would require new code on the server side. It
would make no sense to assign a number twice to the same machine, or
maybe even to the same account. It would also take more
bookkeeping when a number can be running on two clients at once (I
assume).
Finally, it isn't guaranteed to work. It's always possible
that the number will be reassigned to another machine which is just
as slow. The odds of this are low (for some definitions of low :>),
but they are there...
I'm honestly not sure if this is a good idea, but I haven't
seen it mentioned before. The orderly part of me thinks that it's
useful to have as large a contiguous range of checked numbers as
possible, and this plan would clean out some of the low exponents
which will take a huge amount of time if things are left as they
are. The rest of me says that things will catch up eventually, no
need to worry (or someone will do the lowest remaining ranges by
hand, regardless of who has them checked out). Any ideas?
--
Kevin Jaget (or an FDA approved generic equivalent)
kcjaget at mindspring.com
