On 26 Sep 00, at 0:57, FRANK_A_L_I_N_Y wrote:
> I think part of the problem is the current length of time for one LL test
> to complete.
With due respect, what "problem"? For ordinary assignments, the
execution time itself is not a real problem, though there may well be
a psychological problem - it took me a couple of years to build up to
facing runs lasting many months, and I can certainly see that many
people would like to do things that produce real results relatively
quickly, at least during the early stages of participation.
This "problem" may affect recruitment to the project.
For assignments taking many, many months, there is some evidence that
unreliability of current hardware may result in error rates which are
dangerously high. I fear that more frequent crosschecking, or lowered
levels of ambition, may be required.
> It would help if we could parallel process the mega virtual
> machine on one suspect at a time.
The problem here is that the "assignment state vector" (effectively
the save file) is too large to transfer frequently over slow comms
links. Most participants are probably still using analogue modems;
frequently transferring several megabytes across such a link is just
not practical.
The save file can be compressed to something of the order of 75% of
its file size using Zip (or something similar), but this is not
enough to make a real difference. Mathematically it is obvious that
the number of bits in the assignment state vector needed to transfer
an assignment from one host to another must be at least as large as
the exponent being worked on.
There is an alternate approach, which I have mentioned more than once
in the past. The idea is that PrimeNet would assign an exponent to
two users simultaneously; at reasonable intervals, the two users
would crosscheck their work - this would involve only a small amount
of data exchange, since checking the low 64 bits of the residual
would be sufficient. There are obvious problems e.g. PrimeNet would
need extensive redesign; how to deal with "dropouts"; who gets the
kudos if/when a prime is found? Nevertheless I feel that such an
approach may be becoming essential, at least so far as the 10 million
digit exponents are concerned.
I'm actually running a sort of pilot of this involving double-
checking of my residuals on M33219281 up to iteration 17 million,
which I think proves the desirability of such an approach, and also
points out some of the practical problems. Incidentally, I'd be
grateful for a status report from any of the six people who haven't
contacted me since I sent them instructions - even if only to say
you've dropped out.
> Any math majors out there ?
Sure ... but is this relevant in this respect? It's well known that
the FFT can be rearranged for parallel processing, but the problem of
communication between the processors remains. Parallel FFT is
definitely worthwhile on multiple processors sharing main memory
through multiporting, but would probably be impractical even on
clusters of systems connected by local gigabit networks.
Though the speed of networks is bound to increase, it is probable
that the demands of increasing exponent size will more than negate
this - just as the time taken to run a typical assignment now is much
greater than it used to be, despite rapidly increasing CPU power.
> If Transmeta
> takesoff, anyone with an internet appliance or cell phone with web access
> could run prime95. www.transmeta.com
Theoretically, yes.
In practice, if the near-state-of-the-art processors in desktop
systems are slow enough that run times are becoming excessively long,
how long is it going to take to do anything useful in the distinctly
limited power available in appliances designed for intermittent use
with low-power applications?
Don't forget that transmitting the necessary state vectors is going
to take a long time compared with executing a few iterations.
Also, I certainly wouldn't want to pay the call charges involved in
up- & downloading multimegabyte files on a mobile phone on a frequent
basis.
Regards
Brian Beesley
_________________________________________________________________________
Unsubscribe & list info -- http://www.scruz.net/~luke/signup.htm
Mersenne Prime FAQ -- http://www.exu.ilstu.edu/mersenne/faq-mers.txt
