The "(W)TF Depth Debate" state of the bunion - Page 10

c10ck3r · 2013-12-15, 19:29

As much as I like to disagree with Daveiddy whenever possible, I feel like at the minimum I should pose this question:

Would it be beneficial to (at least temporarily) take the 64M-70M exponents only to 73, and re-purpose half the power to taking 50M exponents to 71 or 72? It seems like instead of funneling all of our energy to push the wave forward, perhaps we should ensure that the factor-filled land behind us has been thoroughly scoured of low factors...

Eventually, we could get rid of "DCTF" if we do everything in one fell swoop from now on, but maybe that would require stopping the wave to catch our breath?

Edit:
Oh dear, he'll be the first one to see this post. What hath I wrought?

TheMawn · 2013-12-15, 20:04

I don't know that doing DCTF and TF all in one go is exactly reasonable. That would require us to eventually reach the point where TF in 65M would be done up to 75, which is just not reasonable. The point Davieddy tries to make is that even 74 isn't reasonable.

To be honest I've never given any thought to DCTF because we're so far ahead of that wave. It DOES need to get done eventually, but because LL and TF are so neck and neck, we can't really afford to move anything over. The work I can contribute to the TF front is needed in a week, but the work I can contribute to the DCTF is only needed in maybe six months or a year or whatever (I truly do not know).

On the other hand, my feeling is that there is a serious lack of participation in the DC department. I switched half of my CPU to DC a while back, and at something like 15 DC for 500 GHz-Days and 50 LL for 6500 GHz-Days, my DC and LL ranks were the same. For every 3 LL that get done, a single DC gets done, apparently.

[I went and checked mersenne.info and in the last month, there were 5,300 new "One LL" exponents and 3,000 new "Two LL" exponents, so 1 in 3 is pretty close (3 in 8.3 is more exact)]

If Primenet^R Makes Sense^TM was to emphasize DC's a bit more (say 15% of WMS assignments are DC) to try to bring them up to speed, we could re-purpose some of our TF to DCTF (say 10%) while increasing the lead in the TF front (the remaining 5%).

I think re-balancing the LL/DC emphasis is a better option than changing the factoring limits.

chalsall · 2013-12-15, 20:16

Quote:

Originally Posted by c10ck3r

Would it be beneficial to (at least temporarily) take the 64M-70M exponents only to 73, and re-purpose half the power to taking 50M exponents to 71 or 72?

IMO (and many others), no.

Because it takes longer to LL the higher the candidate, it Makes (more) Sense ^(TM) to eliminate the higher candidates by TFing first.

And because we're only slightly ahead of the LLing "wave", we should focus on what we're currently doing, rather than "dropping back down" into the 50's.

It is only because we have a slight lead that we're continuing to bring in 62M's TFed to "only" 73 to bring up to 74.

Edit: Please don't forget this report. No "50's" are at less than 72.

Prime95 · 2013-12-15, 20:54

Quote:

Originally Posted by chalsall

IMO (and many others), no.

This doesn't pass a simple common sense test. Pose the question this way: Would GIMPS be better off if your GPU took two 58M exponents from 72 to 73 or one 64M exponent from 73 to 74?

1) The 2 58M TFs take only a little more time than the 1 64M TF.
2) The 2 58M TFs have a little more than twice the chance of finding a factor.

I claim the extra iterations and higher FFT length of a saved 64M LL test does not compensate for the nearly double chance of saving one 58M LL test.

chalsall · 2013-12-15, 21:14

Quote:

Originally Posted by Prime95

I claim the extra iterations and higher FFT length of a saved 64M LL test does not compensate for the nearly double chance of saving one 58M LL test.

What are your orders, Admiral?

I'm serious.

As you know, I'm more than able to bring in lower candidates to take to 73. Happy to facilitate if you think this is advisable.

garo · 2013-12-15, 22:05

Geroge has a very good point. Whatever your threshold is for taking everything to 74 - and at the moment it seems that anything over 60M is taken to 74, the threshold for 73 should be ~14M behind. The way things are currently, there is a jump of two bits at 60M.

BTW, how many exponents from 50-60M are made available everyday anyway? I'd vote for capturing those and taking them to 73 assuming we have sufficient TF cycles available. And if we don't I would rather we throw back 60-64Ms at 73.

chalsall · 2013-12-15, 22:17

Quote:

Originally Posted by garo

I'd vote for capturing those and taking them to 73 assuming we have sufficient TF cycles available. And if we don't I would rather we throw back 60-64Ms at 73.

[In the voice of Data]I copy you commander.[/Data]

I don't actually think that makes the most sense, but let's see how that works out; we might actually have the margin to pull it off (but only just)....

TheMawn · 2013-12-16, 02:12

Quote:

Originally Posted by Prime95

This doesn't pass a simple common sense test. Pose the question this way: Would GIMPS be better off if your GPU took two 58M exponents from 72 to 73 or one 64M exponent from 73 to 74?

1) The 2 58M TFs take only a little more time than the 1 64M TF.
2) The 2 58M TFs have a little more than twice the chance of finding a factor.

Ah. By 50M I thought Clocker was referring to DC territory, but the upper range of 50M is actually still LL. I'm working on two of those as we speak...

LaurV · 2013-12-16, 02:35

I have a point for George, and one against him. I mean, not really, in fact I have a point for doing TF to 73 only, but a point against doing TF to higher bits in the 50M range.

The point "for him" would go like "the TF to 64 is not really productive". This bitlevel is at the limit, you get about the same "clearance rate" doing TF to 64 in the current LL range, or doing LL**. This I can say for sure, after I have TF-ing from 71/72 to 74 in the last week (10 days), with about 900GHzDays/day. I only found 9 factors, and Chris and Kracker both found 10 factors each, doing P-1 with about the same rate (edit: both together, not each) I could do P-1 if I would use the cards to do P-1. A little more, in fact, but also two times more factors. And my factors are most of them from below 74 bits, in fact I think there are only 2 or 3 which are "pure" 74 bits factors. I could clear 3 exponents by LL/DC in this time (6 tests on 2 cards take ~12 days in fact, with "normal", factory overclock, without tricks). So, upon your luck, you can get a little bit better doing TF to 74, or a little bit worse. The best way for the current range (and this I am advocating since long time, and I can prove by calculus, and by empirical data from my work and from Chris' tables) is to TF to 73, then do P-1, lots of P-1. Here I come in the sentiment with RDS, which also recommended P-1 against TF (but he is irrationally doing this at every bit level :P). Practically, the fastest way one can eliminate exponents now, at the current LL front, and since cudaPM1 is available, is to do P-1.

The point "against" stands on the fact that George's calculus does not take consideration of the fact that the DC range had a lot of P-1 done already, so the chances of finding factors is much lower there. Assuming you spend x time to do one 64M to 74 bits, and you have y chance to find a factor, then: to TF one 58M to 73, you will need to pass through the half of the quantity of factor candidates, but you crawl through them with a 2*58M step instead of 2*64M step, therefore to TF two of them, you will need 64/58 x time, about 10% more time, but your chance to find a factor is not 2y, it is only about 1.3y, due to the P-1 done in the range. So, you spend about 10% more time, and have about 30% more chance to clear one exponent.

Whatever path is chosen, we will follow with our GPUs.

By the way, something is wrong with this Oliver guy... We put together almost all our P-1 factoring power (Chris, Kracker and me), so now we together are faster than any other participant taken alone, by far (we are only doing this since 10 days, and look to the numbers in the rightmost column), and Oliver is still not in sight... Neither a ">999" either! (which would mean that we are still faster, but the deadline is far away). Nothing!

Something is wrong with this guy, someone stop him!

Click image for larger version

Name: olie.PNG
Views: 82
Size: 57.7 KB
ID: 10575

-------------------
** with fast DP cards like GTX580's and Titans (lower cards like 560, 660, etc, still can do better by doing TF)

kracker · 2013-12-16, 02:50

He has a lot of power...

Hmm... that chart doesn't look exactly right for some reason.

LaurV · 2013-12-16, 02:52

Quote:

Originally Posted by chalsall

[In the voice of Data]I copy you commander.[/Data]

I don't actually think that makes the most sense, but let's see how that works out; we might actually have the margin to pull it off (but only just)....

I added 500GHzDays/day to my battery last night (see this post, after I was able to install .Net 4, and run CCC and Misfit with it, I could set the proper clocks for the card, ran "-st2" and everything was fine, so I am crunching right now with 1400G/day - the "top difference" was 3900G, now is going down

. Technically, I am in the 100%-ile, so you can remove your CPU's now. I will continue to TF, and if you want me to "save" the result files and send them to you instead of reporting it, please PM).

2013-12-15, 19:29	#100
c10ck3r Aug 2010 Kansas 547 Posts	As much as I like to disagree with Daveiddy whenever possible, I feel like at the minimum I should pose this question: Would it be beneficial to (at least temporarily) take the 64M-70M exponents only to 73, and re-purpose half the power to taking 50M exponents to 71 or 72? It seems like instead of funneling all of our energy to push the wave forward, perhaps we should ensure that the factor-filled land behind us has been thoroughly scoured of low factors... Eventually, we could get rid of "DCTF" if we do everything in one fell swoop from now on, but maybe that would require stopping the wave to catch our breath? Edit: Oh dear, he'll be the first one to see this post. What hath I wrought? Last fiddled with by c10ck3r on 2013-12-15 at 19:30 Reason: Teh lolz

2013-12-15, 20:04	#101
TheMawn May 2013 East. Always East. 11×157 Posts	I don't know that doing DCTF and TF all in one go is exactly reasonable. That would require us to eventually reach the point where TF in 65M would be done up to 75, which is just not reasonable. The point Davieddy tries to make is that even 74 isn't reasonable. To be honest I've never given any thought to DCTF because we're so far ahead of that wave. It DOES need to get done eventually, but because LL and TF are so neck and neck, we can't really afford to move anything over. The work I can contribute to the TF front is needed in a week, but the work I can contribute to the DCTF is only needed in maybe six months or a year or whatever (I truly do not know). On the other hand, my feeling is that there is a serious lack of participation in the DC department. I switched half of my CPU to DC a while back, and at something like 15 DC for 500 GHz-Days and 50 LL for 6500 GHz-Days, my DC and LL ranks were the same. For every 3 LL that get done, a single DC gets done, apparently. [I went and checked mersenne.info and in the last month, there were 5,300 new "One LL" exponents and 3,000 new "Two LL" exponents, so 1 in 3 is pretty close (3 in 8.3 is more exact)] If Primenet^R Makes Sense^TM was to emphasize DC's a bit more (say 15% of WMS assignments are DC) to try to bring them up to speed, we could re-purpose some of our TF to DCTF (say 10%) while increasing the lead in the TF front (the remaining 5%). I think re-balancing the LL/DC emphasis is a better option than changing the factoring limits. Last fiddled with by TheMawn on 2013-12-15 at 20:06

2013-12-16, 02:35	#108
LaurV Romulan Interpreter Jun 2011 Thailand 5·1,931 Posts	I have a point for George, and one against him. I mean, not really, in fact I have a point for doing TF to 73 only, but a point against doing TF to higher bits in the 50M range. The point "for him" would go like "the TF to 64 is not really productive". This bitlevel is at the limit, you get about the same "clearance rate" doing TF to 64 in the current LL range, or doing LL. This I can say for sure, after I have TF-ing from 71/72 to 74 in the last week (10 days), with about 900GHzDays/day. I only found 9 factors, and Chris and Kracker both found 10 factors each, doing P-1 with about the same rate (edit: both together, not each) I could do P-1 if I would use the cards to do P-1. A little more, in fact, but also two times more factors. And my factors are most of them from below 74 bits, in fact I think there are only 2 or 3 which are "pure" 74 bits factors. I could clear 3 exponents by LL/DC in this time (6 tests on 2 cards take ~12 days in fact, with "normal", factory overclock, without tricks). So, upon your luck, you can get a little bit better doing TF to 74, or a little bit worse. The best way for the current range (and this I am advocating since long time, and I can prove by calculus, and by empirical data from my work and from Chris' tables) is to TF to 73, then do P-1, lots of P-1*. Here I come in the sentiment with RDS, which also recommended P-1 against TF (but he is irrationally doing this at every bit level :P). Practically, the fastest way one can eliminate exponents now, at the current LL front, and since cudaPM1 is available, is to do P-1. The point "against" stands on the fact that George's calculus does not take consideration of the fact that the DC range had a lot of P-1 done already, so the chances of finding factors is much lower there. Assuming you spend x time to do one 64M to 74 bits, and you have y chance to find a factor, then: to TF one 58M to 73, you will need to pass through the half of the quantity of factor candidates, but you crawl through them with a 258M step instead of 264M step, therefore to TF two of them, you will need 64/58 x time, about 10% more time, but your chance to find a factor is not 2y, it is only about 1.3y, due to the P-1 done in the range. So, you spend about 10% more time, and have about 30% more chance to clear one exponent. Whatever path is chosen, we will follow with our GPUs. By the way, something is wrong with this Oliver guy... We put together almost all our P-1 factoring power (Chris, Kracker and me), so now we together are faster than any other participant taken alone, by far (we are only doing this since 10 days, and look to the numbers in the rightmost column), and Oliver is still not in sight... Neither a ">999" either! (which would mean that we are still faster, but the deadline is far away). Nothing! Something is wrong with this guy, someone stop him! ------------------- * with fast DP cards like GTX580's and Titans (lower cards like 560, 660, etc, still can do better by doing TF) Last fiddled with by LaurV on 2013-12-16 at 02:53

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
Specifing TF factor depth in "Manual Assignments"?	kracker	PrimeNet	2	2012-07-22 17:49
Would Minimizing "iterations between results file" may reveal "is not prime" earlier?	nitai1999	Software	7	2004-08-26 18:12

2013-12-15, 22:05	#105
garo Aug 2002 Termonfeckin, IE 2²·691 Posts	Geroge has a very good point. Whatever your threshold is for taking everything to 74 - and at the moment it seems that anything over 60M is taken to 74, the threshold for 73 should be ~14M behind. The way things are currently, there is a jump of two bits at 60M. BTW, how many exponents from 50-60M are made available everyday anyway? I'd vote for capturing those and taking them to 73 assuming we have sufficient TF cycles available. And if we don't I would rather we throw back 60-64Ms at 73.

2013-12-16, 02:50	#109
kracker "Mr. Meeseeks" Jan 2012 California, USA 4170₈ Posts	He has a lot of power... Hmm... that chart doesn't look exactly right for some reason.