New milestone

[NBtarheel_33]

Originally Posted by Primeinator

Yes. It took nearly 7 years after finding M40 to prove it was truly the 40th Mersenne prime.

This is going to get worse, unless we get some Tennesseans...I mean volunteers...to take up a decent share of double-checking duties. Right now, based on the pace of returned results from May 21, 2010 to date, we are looking at not knowing the true positions of M46 and M47 until late in the year 2020. That would be 11-12 years from date of discovery to proven position on the Mersenne list.

I mentioned in another post that I wonder how many folks would balk if George suspended all first-time LL testing for six months, and handed out only double-checks. I am considering letting all or nearly all of my cores switch over to double-checking exclusively for a few months.

*Another option* - with TF being so overpowered (and GIMPS about to run out of exponents to TF in the classical 79.3M assignment space), maybe we should suspend TF in favor of DC, except for the absolutely oldest and slowest boxen on PrimeNet - which should be getting the TF-LMH assignments that are being needlessly chewed on (nay, ravenously devoured) with state-of-the-art boxen.

[Primeinator]

Originally Posted by NBtarheel_33

I am considering letting all or nearly all of my cores switch over to double-checking exclusively for a few months.

Isn't this like using a 1mL dropper to drain a barrel? It is going to take the work of many individuals to dent the number of double checks required up to M47 (or even close).

[Prime95]

I can think of two viable options:

1) Change the first-time / double-checking crossover. Maybe increase it to 2500 MHz. See http://www.mersenne.org/thresholds/

2) Offer bonus CPU-time credit for double-checking. This goes against GIMPS long history of trying to accurately give CPU credit. I'd bet a 50% bonus would convince some of those TFers to switch to double-checking.

[NBtarheel_33]

Originally Posted by Primeinator

Isn't this like using a 1mL dropper to drain a barrel? It is going to take the work of many individuals to dent the number of double checks required up to M47 (or even close).

Well, let's see, I have 39 cores (plus 5 or 6, I think, that are yet to be borged). Right now, I'm finishing one first-time LL on average every 1.25 days. If I switched this effort to doublechecks, I could probably what - finish a DC every 0.5-0.6 days on average? That would be 600-700 doublechecks in a year. Right now, we're averaging a return of 38-39k doublechecks per year below M43112609. So if I turned my whole borg to DCs, I'd increase annual throughput by 1.5-2% give or take. Not a huge contribution, but nothing to sneeze at either.

On the other hand, I wonder if we could get the heaviest hitters (curtisc, ahmerali, linde, etc.) to switch some or all of their machines over to DC for a while. Not sure how these folks feel about finding a prime vs. placing Mxx in its proper place.

George - changing the threshold might be a good idea. I *would not* mess with the credit system. It's more interesting if it accurately reflects a user's total effort, not some sort of make-believe scaled number (and someone could easily figure out the real number anyway, just by subtracting .50 * the number of GHz-days earned from DCs).

Here's another option, not sure how feasible it is. Presumably there is some prize money in the GIMPS coffers to be given out as research discovery prizes. How about a token prize (say $100 or something) for the highest doublecheck throughput over some appropriate interval. For example, we could advertise a doublecheck drive for the year beginning September 1, 2010. Whoever has the best throughput (measured by some appropriate metric, say total GHz-days divided by total computing power (to make it fair whether you have one computer or 100)) for DCs from 9/1/10-9/1/11 gets the $100 prize (and recognition on the GIMPS homepage).

[Primeinator]

Originally Posted by NBtarheel_33

Well, let's see, I have 39 cores (plus 5 or 6, I think, that are yet to be borged). Right now, I'm finishing one first-time LL on average every 1.25 days. If I switched this effort to doublechecks, I could probably what - finish a DC every 0.5-0.6 days on average? That would be 600-700 doublechecks in a year. Right now, we're averaging a return of 38-39k doublechecks per year below M43112609. So if I turned my whole borg to DCs, I'd increase annual throughput by 1.5-2% give or take. Not a huge contribution, but nothing to sneeze at either.

My apologies! That is far more than removing drops at a time. I retract my previous statement.

I would be curious to know if not only some of the heavy hitters, but some of the individuals working on other projects with higher success rates (arguably more practical) would be willing to contribute to the DCing process? After all, in so doing they are achieving a more tangible result than first time LLs and are doing a great service by putting M41- M47 in proper numerical order.

[markr]

Originally Posted by Prime95

I can think of two viable options:

1) Change the first-time / double-checking crossover. Maybe increase it to 2500 MHz. See http://www.mersenne.org/thresholds/

2) Offer bonus CPU-time credit for double-checking. This goes against GIMPS long history of trying to accurately give CPU credit. I'd bet a 50% bonus would convince some of those TFers to switch to double-checking.

I'd also prefer the credit system reflected effort alone - please don't add arbitrary factors to it. Changing the crossover would be better. (For anyone who hasn't checked the thresholds page, it's currently 2000 MHz, but that's a "Pentium 4 equivalent speed" so it's not as high as it sounds.)

What if all the "whatever makes sense" machines above that threshold were assigned first-time LLs and double-checks alternately, or randomly in some proportion, or even just always assign one or two to any new machine? IIRC you considered a similar idea in the "P-1 factoring anyone?" thread - to assign some P-1 as long as the GIMPS client was configured with sufficient memory. A few DCs for each PC would also help measure their reliability.

[garo]

I agree with markr above. What if every new machine was assigned a doublecheck first unless the user explicitly requests not to?

[NBtarheel_33]

Curtisc have handed in 7,092 LLs in the last 365 days. A conservative estimate is that a DC takes half the time of a current LL (for now, anyway, as we'll be working in the 20Ms and low 30Ms to start). So, figure that curtisc could perform roughly 14,000 doublechecks in a year if they were to switch their computers over to DC. This alone would increase doublecheck throughput by 50% or so on an annual basis, bringing the estimated wait for a proof that M47 is really M47, down to around 7 years from its date of discovery (i.e. 2016 or so).

[garo]

curtisc is in it to find more primes. I don't think it is fair to ask people like him to do DCs. Instead we should get all new entrants - most of whom never finish an LL because it takes so bloody long - to do one or more DCs first instead.

[petrw1]

Originally Posted by Prime95

I can think of two viable options:

1) Change the first-time / double-checking crossover. Maybe increase it to 2500 MHz. See http://www.mersenne.org/thresholds/

2) Offer bonus CPU-time credit for double-checking. This goes against GIMPS long history of trying to accurately give CPU credit. I'd bet a 50% bonus would convince some of those TFers to switch to double-checking.

I too would prefer to NOT have a bonus system.

This is similar to the problem last year (and maybe still a problem) of not having enough doing large P-1. I'd be curious to know how many PCs (as a percentage) are set to "Whatever makes sense". i.e. are there enough that you could use them to make bigger gain in P-1 and DC?

[henryzz]

Originally Posted by garo

curtisc is in it to find more primes. I don't think it is fair to ask people like him to do DCs. Instead we should get all new entrants - most of whom never finish an LL because it takes so bloody long - to do one or more DCs first instead.

Agreed. If you change the threshold then please make sure that Q6600s fit first-time tests. It really is getting to something if desktop quads are being handed DCs because they are too slow. I can understand mobile quads as they are clocked low.