New work types for GPU72?

[NormanRKN]

Originally Posted by chalsall

5. GPU72 is simply a coordination tool for those with GPUs to do the newly appropriate TFing without "stepping on each other's toes" (read: two or more people doing the same work on the same candidate).

how can that happen that a few people doing the same work at the same time ? are the candidates not blocked/reserved from mersenne.org for a specific user or you mean instead more the work of "other" sources ?

Norman

[chalsall]

Originally Posted by NormanRKN

how can that happen that a few people doing the same work at the same time ? are the candidates not blocked/reserved from mersenne.org for a specific user or you mean instead more the work of "other" sources ?

It is possible for people to do and submit work without having the candidate reserved from PrimeNet. This was quite common for a few of us who were doing shallow and broad TFing above 100M a couple of years ago.

Another thing the GPU72 tool does is automates the process of determining which candidates are not currently being worked by anyone else, including (for example) LLing or P-1ing, and makes them available for the appropriate efforts. If you see mention of "Spidy" in the GPU72 sub-forum threads, this is the automatic spider(s) which does this so humans don't have to.

[NormanRKN]

ah ! ok !
another thing.
in gpu72 i can choose factors up to 73. i can set a little higher to 74 or 75 but it appears a "warning" todo that not.
is it bad if i do some work of this or i shouldn't do that?

Norman

[chalsall]

Originally Posted by NormanRKN

another thing. in gpu72 i can choose factors up to 73. i can set a little higher to 74 or 75 but it appears a "warning" todo that not. is it bad if i do some work of this or i shouldn't do that?

The policy of GPU72 (and the Mersenne.org community in general) is that people should be allowed to do any kind of work they want, so long as they don't "poach" (do work that someone else is doing), or "hoard" (reserve more candidates than they can complete in a reasonable amount of time).

The reason the form warns you about going above the recommended level (currently 73) is it doesn't really make sense to do so. So you know, it takes three times as much effort (read: processing time) to TF from (for example) 72 to 74 as it does to TF from 72 to 73 for any given candidate. Also, while it takes three times as long, your chances of finding a factor only increases by approximately two.

So, rather than doing one candidate from 72 to 74, it makes more sense to take three from 72 to 73.

[Dubslow]

Originally Posted by chalsall

The reason the form warns you about going above the recommended level (currently 73) is it doesn't really make sense to do so. So you know, it takes three times as much effort (read: processing time) to TF from (for example) 72 to 74 as it does to TF from 72 to 73 for any given candidate. Also, while it takes three times as long, your chances of finding a factor only increases by approximately two.

So, rather than doing one candidate from 72 to 74, it makes more sense to take three from 72 to 73.

This sort of tripling of work from a to a+2 instead of a to a+1 is true for any (positive integer) a besides a==72; the difference is the decreased chance of a factor at each bit level.

Up to 73, you're likely to find approximately one factor among many exponents for about the same effort as two LL tests on one exponent. If you go beyond 73 bits, then you'll find less than one factor per two-LL's worth of effort, so you will "clear" more exponents by just running the LL test instead of doing more trial factoring.

Despite this, some people aesthetically prefer finding a factor instead of "merely" two matching composite LL tests. That's why the option is available, but it isn't as efficient for the work done.

[LaurV]

As one image is better then a thousand words, look carefully to this green-orange-cyan-red table. The effort for LL starts high and grows higher SLOWLY with the growth of the exponent. The effort to TF starts low, and grows FASTER as the bitlevel grows (it doubles with each bitlevel). You can clear a given exponent either by finding a factor, or by doing two LL. To find a factor you can try to TF higher and higher, but if no factor is found, then you must stop somewhere (at the cyan line), because behind of that point doing TF is more costly (in time) than doing two LLs.

[NormanRKN]

mersenne and gpu72 only give work for TF that make sense (per default) ?
if so we can cool down and relax. ;)

[chalsall]

Originally Posted by NormanRKN

mersenne and gpu72 only give work for TF that make sense (per default)? if so we can cool down and relax. ;)

Yup. That's the whole point.

[NormanRKN]

and the chance to find factors is ~ ???

[James Heinrich]

Originally Posted by NormanRKN

and the chance to find factors is ~ ???

Approximately 1% per trial-factoring bit level (e.g. 2⁷¹-2⁷²), approximately 3-5% for P-1 (depending on the bounds selected and how much trial factoring has been done before).

edit: you can get more specific probability numbers here:
http://mersenne-aries.sili.net/prob.php -- P-1 probability
http://mersenne-aries.sili.net/credit.php -- TF probability

[chalsall]

Originally Posted by chalsall

OK, this work type [DC P-1] has now been made available from GPU72.

Hey All. Just to let you know...

Thanks mostly to the work of diamonddave, assisted by Pete, Wayne and a few others, this sub-sub-project is now complete after approximately 15 months of work.

228 factors found by way of DC P-1'ing.