Will Any Current 100M Digit LL Tests Finish?

[davieddy]

Originally Posted by retina

If you mean an exponent of 100M then the answer is yes. It was StarQwest that did 100,000,007. See http://mersenneforum.org/showthread.php?t=6389 for some details.

Yes and double checked!
But 100M decimal digits (which is what this thread is talking about)
has an exponent of 332M digits and the time needed goes up more
than the exponent^2.

[mdettweiler]

Originally Posted by retina

If you mean an exponent of 100M then the answer is yes. It was StarQwest that did 100,000,007. See http://mersenneforum.org/showthread.php?t=6389 for some details.

Hmm...from the context of this thread, I would guess that dominicanpapi87 meant to say "100M digit tests", not 100M--though, of course, given what he actually *did* post, then your answer is of course correct.

As for 100M digit tests (~333M, if memory serves?), none has yet finished; StarQwest started on one a while back, but seems to have paused on it. (Not sure if he's still submitting progress or not.)

Edit: Ah, I see davieddy beat me to the punch on this.

[JuanTutors]

Yes, by "100M Mersenne" I meant "100M digit Mersenne". Sorry, I thought that was clear.

So really, no one has done one yet? That's really surprising, despite the enormous amount of time it takes. I figured someone would have gotten started on one long enough ago to have finished one by now.

Maybe I should start a new thread asking how far along people are on their 100M digit test :) I hope I finish the first!!

[James Heinrich]

Originally Posted by dominicanpapi82

So really, no one has done one yet? That's really surprising, despite the enormous amount of time it takes.

That's not that surprising to me. On an overclocked (3.5GHz) Q6600 (quad-core), which (even though mine's over a year old now) is at the upper end of consumer hardware, even running on 4 threads a LL test will take somewhere between 2.5 and 3 years at best (not counting TF and P-1). Consumer hardware this powerful wasn't available 3 years ago, so if it's still running on the original machine the test time would be much longer. I'd be surprised if any 100M-digit tests are more than 50% done, even if someone started it some time (1+ years) ago.

Granted, if someone put their mind to it and had access to a supersystem (such as the ones that are used to verify reported primes) a 100M-digit exponent could probably have been tested by now, but it's probably not a "worthy" enough goal to spend months(?) of expensive computer time on.

It seems to me that assignments that take many years are not reasonable to attempt: beyond the risk of the test or machine being lost/corrupted over that span of time, by the time you've spent 4 of 5 years completing the test, then-current hardware could complete the same test in 6 months and still finish the test faster than your computer that's been running for 4 years already. I'm currently working on M332203901 and there's roughly 2 months each of TF and P-1 going into the assignment before LL testing even starts, which would then take another 3 years or so (current expected completion date is 21-Dec-2011 on 4 threads, a lot longer on fewer threads). But by mid-2011, I fully expect to be able to load it up on my Phenom-IV 12-core machine and be done in a few months, which is sooner than if I just left this machine churning for the next 1000 days. (In case you hadn't guessed, I don't plan to proceed to LL testing my exponent, I'll release it once I've finished attempting to factor it). Good luck on being the first to finish

[JuanTutors]

Originally Posted by James Heinrich

Good luck on being the first to finish

Thanks :) I actually do intend on finishing mine. I just don't think computing power with multiple processors will still be able to move data around enough to be able to finish a 100M digit Mersenne that quickly by the time I'm almost done with this one.

Hopefully I won't get tired of it by then!! I was actually thinking of starting some DC or factoring stuff, just to see SOME results along the way.

[retina]

Originally Posted by dominicanpapi82

I just don't think computing power with multiple processors will still be able to move data around enough to be able to finish a 100M digit Mersenne that quickly by the time I'm almost done with this one.

I think it would be interesting to calculate the required data bus bandwidth needed to do a 6 month 100Mdigit LL test. Assuming perfect synchronisation within the cores and never an idle memory bus cycle I wonder what sort of data rate is required. Very high bandwidths would probably need multiple channels of RAM or ultra wide data bus widths or both.

[lpmurray]

I have been working on a hundred million digit number for about 8 months on and off and have reached 11% and have a completion date of 09/2012

[CRGreathouse]

Originally Posted by James Heinrich

That's not that surprising to me. On an overclocked (3.5GHz) Q6600 (quad-core), which (even though mine's over a year old now) is at the upper end of consumer hardware, even running on 4 threads a LL test will take somewhere between 2.5 and 3 years at best (not counting TF and P-1). Consumer hardware this powerful wasn't available 3 years ago, so if it's still running on the original machine the test time would be much longer. I'd be surprised if any 100M-digit tests are more than 50% done, even if someone started it some time (1+ years) ago.

With a simple model where only one computer can be purchased (no upgrading or anything), and where the computers offered for sale increase in performance smoothly by Moore's law (double every 18 months), it's not hard to numerically solve the nonlinear equation to find the ideal time to start a job in order to have the solution as rapidly as possible. A job that today takes one year should be started 20 months ago; if so, it will finish in about 6 months (26 months total). A job that takes 100 years now should be started in 99.5 months; it will finish 125.5 months from now (26 months calculating, 99.5 months waiting).

Under this model, any job taking less than 26 months (2.164 years) should be started now; any job taking more than 26 months should be delayed until it would take only 26 months on current hardware. Insofar as this model reflects reality and the tests currently take 3 years, the first jobs to finish would be the ones started in August 2009, which will finish around October 2011.

[starrynte]

Originally Posted by CRGreathouse

Under this model, any job taking less than 26 months (2.164 years) should be started now; any job taking more than 26 months should be delayed until it would take only 26 months on current hardware.

I'm sure there is a reason, but I can't quite find out how the number 26 is found...

[retina]

Originally Posted by CRGreathouse

With a simple model where only one computer can be purchased (no upgrading or anything), and where the computers offered for sale increase in performance smoothly by Moore's law (double every 18 months), it's not hard to numerically solve the nonlinear equation to find the ideal time to start a job in order to have the solution as rapidly as possible. A job that today takes one year should be started 20 months ago; if so, it will finish in about 6 months (26 months total). A job that takes 100 years now should be started in 99.5 months; it will finish 125.5 months from now (26 months calculating, 99.5 months waiting).

Under this model, any job taking less than 26 months (2.164 years) should be started now; any job taking more than 26 months should be delayed until it would take only 26 months on current hardware. Insofar as this model reflects reality and the tests currently take 3 years, the first jobs to finish would be the ones started in August 2009, which will finish around October 2011.

No. You should always start your job now. Because when new hardware comes along you can transfer your work to the new machine and continue the job.

If you do nothing now then the job will take longer later, even though the new hardware may be faster you still lose the initial computing time.

[CRGreathouse]

Originally Posted by retina

No. You should always start your job now. Because when new hardware comes along you can transfer your work to the new machine and continue the job.

I explained the model I was working with in my post, which I think is probably a realistic model in that respect. I think that only dedicated prime searchers transfer their old GIMPS directories and work-in-progress files to their new computers.

Also, if data can be transferred for free, there's nothing to analyze -- the optimal starting time is . Perhaps a one-upgrade model would be interesting, though.