Possible extention to the "GPU to 72 Tool" project? - Page 20

lycorn · 2011-11-23, 10:39

Quote:

Originally Posted by chalsall

(...) right now I'm working on something else.

Wow... another brick in the wall!

chalsall · 2011-11-23, 14:29

OK everyone, the GPU to 72 Tool is now ready to automatically assign low LL candidates to PrimeNet workers. Please see http://gpu.mersenne.info/account/getassignments/ll/.

Each candidate has been trial factored to at least 72 "bits" by GPUs, and had P-1 work completed. There are currently 67 available in the 45M range, and 33 in the 46M.

The system is currently retaining the lowest 100 for this service, and will be topped up as GPU and P-1 workers continue their completions. This number will be adjusted over time as we get a feel for how much fire power this work type will have brought to bear.

Note that this is intended for those LL workers who are "trusted" to complete the work in a reasonable time, and thus access must be authorized. Most GPU to 72 Workers already have access, but if anyone is not yet in the system but would like access, please sign up and then PM me here.

The DC version of this will be available later today.

petrw1 · 2011-11-23, 14:42

Quote:

Originally Posted by chalsall

Note that this is intended for those LL workers who are "trusted" to complete the work in a reasonable time

Are you using George's method ... ie if your CPU's "Reliability and Confidence" are at a certain level..Per the Assignment Rules page they need to be 0.9 and 2.

chalsall · 2011-11-23, 14:52

Quote:

Originally Posted by petrw1

Are you using George's method ... ie if your CPU's "Reliability and Confidence" are at a certain level..Per the Assignment Rules page they need to be 0.9 and 2.

No, I don't have access to that data. (I don't think -- if there's a publicly available page on PrimeNet for all users, please let me know.)

Instead I'm going by the amount of work someone has done over the last year.

BTW, the Low Double Check work type is now available as well.

petrw1 · 2011-11-23, 15:07

Quote:

Originally Posted by chalsall

No, I don't have access to that data. (I don't think -- if there's a publicly available page on PrimeNet for all users, please let me know.)

Instead I'm going by the amount of work someone has done over the last year.

BTW, the Low Double Check work type is now available as well.

Not that I know of.

My current plan is to do GPUto72 P-1 till about year end...I should complete close to 500. And then I think I will move to this DC.

James Heinrich · 2011-11-23, 15:13

Quote:

Originally Posted by chalsall

the Low Double Check work type is now available as well.

I've grabbed one to see how it works. ETC: 01-Dec-2011

LaurV · 2011-11-23, 16:08

Our last discussions were quite fruitful. Following the argue here on the forum, I took 60 exponents from DC front and crunched them today for a couple of hours on both GPU's from 67/68 to 69 bits. It was one of that lucky days:

Code:

M25300493 has a factor: 482498669664986964257
found 1 factor(s) for M25300493 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32]

M25381229 has a factor: 444817498910828872201
found 1 factor(s) for M25381229 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32]

M25386671 has a factor: 257812533279578085649
found 1 factor(s) for M25386671 from 2^67 to 2^68 [mfaktc 0.17-Win barrett79_mul32]

I also saw that the "score" decreased within the last 36 hours from "2.4 days per exponent cleared (=factor found)" to "1.9 days per expo cleared" (current score on the overall system progress is "197.3 GHz-days per factor). That is very good, but I am still not convinced, the score should be half of that (i.e. the time should be half per factor, or the amount of factors found in this same time should be double) to really worth TF-ing at DC front, comparing with the time CudaLucas takes to clear an exponent in 26M range (24 hours).

Now as DC work is available, I will switch to that type of work after my current bunch of exponents (137 in the queue, plus 10 at LL front to 72, but that type of work I will continue, as each factor saves 2 LL tests, as discussed earlier) will be all finished. I am for sure "trusted producer", being gimpster from the beginning, but here I still could have a problem if the exponents were first time LL-ed by a third party application: the server will not accept my results. So I will have to manually walk through the Primenet's report to check if the first LL was done by P5 or by other program, for each exponent I would get from GPU272, before starting any work on it, and in case not, to un-reserve it back. And as I know, this is more like a guessing now, like check the uppercase/lowercase of the hex digits, scratch your head, turn around, do whatever else.... then take the risk...

Can anything be done in this direction? Like "assignments for DC work which had first LL done by P95 only"? (here maybe George could help to publish the "shift amount" for the LL's for each exponents? all CudaLucas tests have the shift zero, all P95 work have a random shift which is not zero). And have "assignments for DC work which had first LL done by third parties", which therefore could by double checked with P95 only, for who wants to run DC's on CPU.

Is it possible? Or I am dreaming...

lycorn · 2011-11-23, 19:15

Why don´t you keep the GPUs for TF (that´s where they truly excel) and do the DC work on CPU cores (with Prime95)?
Or, if you really want to use GPUs for LL testing, do first-time tests on them.
Both alternatives would save you (or someone else...) the trouble of analyzing the record of previous LL tests performed.

chalsall · 2011-11-23, 19:55

Quote:

Originally Posted by LaurV

Can anything be done in this direction? Like "assignments for DC work which had first LL done by P95 only"? (here maybe George could help to publish the "shift amount" for the LL's for each exponents? all CudaLucas tests have the shift zero, all P95 work have a random shift which is not zero). And have "assignments for DC work which had first LL done by third parties", which therefore could by double checked with P95 only, for who wants to run DC's on CPU.

Is it possible? Or I am dreaming...

It is possible, but I'm not going to do it...

To do what you ask would involve spidering a separate page for each exponent being returned to PrimeNet. Including those I grab and then immediately return because they're not "interesting" enough to have GPU/P-1 work done on them. This would put significant additional load on the PrimeNet server for very (very) little net benefit.

I would advise you take lycorn advice, and use your GPUs for TFing (preferably in the LL range, because I agree with you DC TF work to 69 is only marginally efficient), and any left over CPU cycles you apply to real DCing.

bcp19 · 2011-11-23, 19:56

Quote:

Originally Posted by LaurV

Our last discussions were quite fruitful. Following the argue here on the forum, I took 60 exponents from DC front and crunched them today for a couple of hours on both GPU's from 67/68 to 69 bits. It was one of that lucky days:

Code:

M25300493 has a factor: 482498669664986964257
found 1 factor(s) for M25300493 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32]
 
M25381229 has a factor: 444817498910828872201
found 1 factor(s) for M25381229 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32]
 
M25386671 has a factor: 257812533279578085649
found 1 factor(s) for M25386671 from 2^67 to 2^68 [mfaktc 0.17-Win barrett79_mul32]

I also saw that the "score" decreased within the last 36 hours from "2.4 days per exponent cleared (=factor found)" to "1.9 days per expo cleared" (current score on the overall system progress is "197.3 GHz-days per factor). That is very good, but I am still not convinced, the score should be half of that (i.e. the time should be half per factor, or the amount of factors found in this same time should be double) to really worth TF-ing at DC front, comparing with the time CudaLucas takes to clear an exponent in 26M range (24 hours).

You need to remember not all of us have a GPU like yours. On my 560 I can complete a DC in 2.4 days where my 450 can do one in 4.3 and I'm guessing the 560 Ti to be around 2.1. Compared to that, my i5 2400 can do a DC in a little under 4 days while my i7 Q740 takes 10 days.

In terms of GHz Days/Day, doing TF, my 560 does ~140, the 460 ~95 and the 560 Ti doing ~150. ~385 GHzDays/Day at your 197.3/factor means I'd average ~2 factors per day compared to completing ~5 DC/4 days. 5 vs 8 tells me that TF outweighs LL at this point in terms of GPU usage on my systems. I would think your GPU to be similiar in terms of factors vs DC.

diamonddave · 2011-11-23, 19:58

Quote:

Originally Posted by lycorn

Why don´t you keep the GPUs for TF (that´s where they truly excel) and do the DC work on CPU cores (with Prime95)?
Or, if you really want to use GPUs for LL testing, do first-time tests on them.

The GPU to 72 project cater in big part to people having that kind of hardware (GPU).

If I like to TF number or do LL, DC or god forbid do OBD or further TF on already DCed exponent. I think, those who pay for the hardware and electricity should have the right to do as they please?

Coming here and asking for a legitimate feature, given the problem encountered by the user. But being rebuked with a, "why don't you do something else" is not very helpful.

Software change and get better... Why wouldn't CUDALucas get so much better that doing any LL on GPU wouldn't be wasteful.

The issue of identifying where the first LL result came from is still relevant.

Quote:

Originally Posted by lycorn

Both alternatives would save you (or someone else...) the trouble of analyzing the record of previous LL tests performed.

Please let him or them determine what is too much trouble

2011-11-23, 14:29	#211
chalsall If I May "Chris Halsall" Sep 2002 Barbados 23047₈ Posts	GPU to 72 now ready to assign real LL work. OK everyone, the GPU to 72 Tool is now ready to automatically assign low LL candidates to PrimeNet workers. Please see http://gpu.mersenne.info/account/getassignments/ll/. Each candidate has been trial factored to at least 72 "bits" by GPUs, and had P-1 work completed. There are currently 67 available in the 45M range, and 33 in the 46M. The system is currently retaining the lowest 100 for this service, and will be topped up as GPU and P-1 workers continue their completions. This number will be adjusted over time as we get a feel for how much fire power this work type will have brought to bear. Note that this is intended for those LL workers who are "trusted" to complete the work in a reasonable time, and thus access must be authorized. Most GPU to 72 Workers already have access, but if anyone is not yet in the system but would like access, please sign up and then PM me here. The DC version of this will be available later today.

2011-11-23, 16:08	#216
LaurV Romulan Interpreter Jun 2011 Thailand 2×3×1,609 Posts	Our last discussions were quite fruitful. Following the argue here on the forum, I took 60 exponents from DC front and crunched them today for a couple of hours on both GPU's from 67/68 to 69 bits. It was one of that lucky days: Code: M25300493 has a factor: 482498669664986964257 found 1 factor(s) for M25300493 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32] M25381229 has a factor: 444817498910828872201 found 1 factor(s) for M25381229 from 2^68 to 2^69 [mfaktc 0.17-Win barrett79_mul32] M25386671 has a factor: 257812533279578085649 found 1 factor(s) for M25386671 from 2^67 to 2^68 [mfaktc 0.17-Win barrett79_mul32] I also saw that the "score" decreased within the last 36 hours from "2.4 days per exponent cleared (=factor found)" to "1.9 days per expo cleared" (current score on the overall system progress is "197.3 GHz-days per factor). That is very good, but I am still not convinced, the score should be half of that (i.e. the time should be half per factor, or the amount of factors found in this same time should be double) to really worth TF-ing at DC front, comparing with the time CudaLucas takes to clear an exponent in 26M range (24 hours). Now as DC work is available, I will switch to that type of work after my current bunch of exponents (137 in the queue, plus 10 at LL front to 72, but that type of work I will continue, as each factor saves 2 LL tests, as discussed earlier) will be all finished. I am for sure "trusted producer", being gimpster from the beginning, but here I still could have a problem if the exponents were first time LL-ed by a third party application: the server will not accept my results. So I will have to manually walk through the Primenet's report to check if the first LL was done by P5 or by other program, for each exponent I would get from GPU272, before starting any work on it, and in case not, to un-reserve it back. And as I know, this is more like a guessing now, like check the uppercase/lowercase of the hex digits, scratch your head, turn around, do whatever else.... then take the risk... Can anything be done in this direction? Like "assignments for DC work which had first LL done by P95 only"? (here maybe George could help to publish the "shift amount" for the LL's for each exponents? all CudaLucas tests have the shift zero, all P95 work have a random shift which is not zero). And have "assignments for DC work which had first LL done by third parties", which therefore could by double checked with P95 only, for who wants to run DC's on CPU. Is it possible? Or I am dreaming... Last fiddled with by LaurV on 2011-11-23 at 16:15

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Stockfish game: "Move 8 poll", not "move 3.14159 discussion"	MooMoo2	Other Chess Games	5	2016-10-22 01:55
Aouessare-El Haddouchi-Essaaidi "test": "if Mp has no factor, it is prime!"	wildrabbitt	Miscellaneous Math	11	2015-03-06 08:17
Collaborative mathematics: the "polymath" project	Dougy	Math	11	2009-10-21 10:04
Would Minimizing "iterations between results file" may reveal "is not prime" earlier?	nitai1999	Software	7	2004-08-26 18:12

2011-11-23, 19:15	#217
lycorn "GIMFS" Sep 2002 Oeiras, Portugal 5C1₁₆ Posts	Why don´t you keep the GPUs for TF (that´s where they truly excel) and do the DC work on CPU cores (with Prime95)? Or, if you really want to use GPUs for LL testing, do first-time tests on them. Both alternatives would save you (or someone else...) the trouble of analyzing the record of previous LL tests performed.