P-1 Question - Page 3

Dubslow · 2012-01-07, 06:47

Quote:

Originally Posted by cheesehead

Stage 1-only is a proper and optimum way to do P-1 on some systems with small allocated memory. It is not "half-assed". It is what was optimum for the conditions in which it was run. You need to understand that.

True statement.

Quote:

Stage 1-only is a proper and optimum way to do P-1 on some systems with small allocated memory. It is not "half-assed". It is what was optimum for GIMPS.

I do not believe that, and judging by some of the tools on his website, I would hazard a guess that James does not either. If others contest this though, I will bow to their better judgement.

cheesehead · 2012-01-07, 06:59

Quote:

Originally Posted by Dubslow

I do not believe that, and judging by some of the tools on his website, I would hazard a guess that James does not either.

I wrote "It is what was optimum for the conditions in which it was run."

Your disbelief is based on your ignorance of the P-1 bounds-choosing algorithm in prime95. It uses parameters that James's page ignores or assumes.

It almost seems that you are treating James's page's approximation as more authoritative than the actual full bounds-choosing algorithm in prime95.

Have you even bothered to study the source code for the bounds-choosing algorithm in prime95, as I have? If not, why do you disbelieve what I tell you about it, when your experience is only with a Webpage approximation, not the real algorithm?

Quote:

If others contest this though, I will bow to their better judgement.

But not to mine, eh? Because you don't like what I'm saying?

Maybe you should ask some other veterans whether I know what I'm talking about when I write about P-1 in GIMPS. I've been here since the first year of GIMPS, before P-1 was even implemented in prime95. I'm not the ultimate expert on P-1, but I know it a lot better than you do, given the number of incorrect statements based on ignorance that you're posting.

axn · 2012-01-07, 07:46

Quote:

Originally Posted by cheesehead

But that's not what the prime95 algorithm does. It chooses B1/B2 such that

(L-L cost) * (factor chance at a particular B1/B2) = (P-1 cost at that same B1/B2

I'll let you figure out why this statement is not entirely correct.

Dubslow · 2012-01-07, 15:19

Quote:

Originally Posted by cheesehead

I wrote "It is what was optimum for the conditions in which it was run."

Please reread my post. I am aware that the bounds choosing algorithm chooses the best bounds possible under the circumstances, but my point is that we can improve the circumstances and do better for GIMPS.

cheesehead · 2012-01-07, 15:41

Quote:

Originally Posted by axn

Quote:

Originally Posted by cheesehead

But that's not what the prime95 algorithm does. It chooses B1/B2 such that

(L-L cost) * (factor chance at a particular B1/B2) = (P-1 cost at that same B1/B2)

I'll let you figure out why this statement is not entirely correct.

I was aware that it is not entirely correct, but I deemed exactitude not to be necessary for the purpose and context for which I was writing it. Consider the equals sign to be replaced by a greater than or equal sign.

cheesehead · 2012-01-07, 16:29

Dubslow,

I apologize for letting my irritation at your belittling references to previous stage 1-only runs interfere with clarity of my explanation.

Quote:

Originally Posted by Dubslow

Please reread my post.

The way you substituted your own wording for mine in the second quote block makes it look like you are quoting me there as well as in the first quote block.

Don't do that.

You should have, instead, written your version without using a quote block, to make it clearer that you were proposing the alteration you presented.

Quote:

I am aware that the bounds choosing algorithm chooses the best bounds possible under the circumstances, but my point is that we can improve the circumstances and do better for GIMPS.

You don't make that point very well with your demeaning characterizations of stage 1-only runs.

The consistent errors in your comments about P-1 bounds, and belittling of stage 1-only P-1, demonstrate that you do not understand the role of allocated memory in the actual bounds-choosing algorithm, probably because James's page does not mention it.

James's page is okay as far as it goes (except it needs to mention its assumptions about allocated memory and timing so as not to mislead newbies like youi). Note in particular that it does not falsely represent itself as (1) a means of learning everything about the actual prime95 bounds-choosing algorithm or (2) an explanation of why there are so many stage 1-only runs among actual P-1 reports.

Regarding the presence of so many stage 1-only results in the database: Why do you think there are so many? (Or have you ever pondered that?)

Answer: because so many systems that ran P-1 had allocated memory no larger than the default value of 8M.

When the prime95 P-1 bounds-choosing algorithm is working on a system that has no more than 8M allocated memory, it is indeed the best and optimum choice from an overall GIMPS standpoint to specify B1=B2 (i.e., stage 1 only) for that particular run!

The only alternative, since stage 2 cannot be run without having more memory for allocating workareas, would be not to do P-1 at all -- clearly not an optimal choice for GIMPS.

Stage 2 must have enough memory for several workareas it will use. Without that, a stage 1-only run is the only feasible alternative. Thus, stage 1-only runs with bounds chosen by the prime95 bounds-choosing algorithm are not "half-assed", but instead are GIMPS's best choice in those situations.

(Note: I don't recall for 100% sure, right now without going back to examine the source code, whether it is absolutely impossible to run a crippled stage 2 without allocated workareas. If it is, then at least such a workarea-less stage 2 would provide not nearly as much advantage over simply using higher B1 during stage 1 as a stage 2 with workareas would.)

Now, it is true that when a system with a generous amount of allocated memory for stage 2 is available, it can do a P-1 run with a higher chance of finding a factor than the earlier stage 1-only run had. But that's only due to its having more memory than the first system did. Not having a lot of allocated memory was not a sin of the first system; it was merely a circumstance, not something to scorn.

Extending a P-1 run to higher B2 (or B1 !) may indeed be good for GIMPS (but note my example showing that it may not be the most efficient use of resources for the advancement of GIMPS). That's fine. But stop the "proper"/"half-assed" characterizations, because they imply that stage 1-only P-1 runs were the result of some poor decision.

(Note: There are some P-1 runs in the database for which the user, instead of letting the prime95 bounds-choosing algorithm specify B1/B2, chose to specify B1/B2 explicitly. But in my present comments, I am referring to runs for which the bounds-choosing algorithm specified B1/B2, unless otherwise noted.)

chalsall · 2012-01-07, 18:23

Quote:

Originally Posted by cheesehead

Extending a P-1 run to higher B2 (or B1 !) may indeed be good for GIMPS (but note my example showing that it may not be the most efficient use of resources for the advancement of GIMPS). That's fine. But stop the "proper"/"half-assed" characterizations, because they imply that stage 1-only P-1 runs were the result of some poor decision.

(Note: There are some P-1 runs in the database for which the user, instead of letting the prime95 bounds-choosing algorithm specify B1/B2, chose to specify B1/B2 explicitly. But in my present comments, I am referring to runs for which the bounds-choosing algorithm specified B1/B2, unless otherwise noted.)

Trying to bring this discussion back to the question at hand (at least for me)...

cheesehead... based on your superior knowledge of P-1 work (at least, superior to mine), what do you think should be the parameters for the (possibly) new work type "P-1 Stage 1 and 2"? I assume it would be a function of the TF level and B1 done (where B1 == B2), but I don't understand the characteristics well enough to design an intelligent assignment algorithm.

The following is a query against the G72 database, sorted by B1 and limited to 50 rows. Please note this is simply what Spidy has seen; it is not a complete analysis from GIMPS.

Code:

+----------+--------+--------+--------+
| Exponent | FactTo | B1     | B2     |
+----------+--------+--------+--------+
| 55204091 |     76 | 405000 | 405000 | 
| 55098229 |     75 | 460000 | 460000 | 
| 55197041 |     75 | 460000 | 460000 | 
| 55835839 |     75 | 470000 | 470000 | 
| 45571601 |     72 | 510000 | 510000 | 
| 46130221 |     72 | 515000 | 515000 | 
| 46207093 |     72 | 520000 | 520000 | 
| 47324509 |     72 | 525000 | 525000 | 
| 54258847 |     74 | 525000 | 525000 | 
| 54476567 |     74 | 525000 | 525000 | 
| 54476573 |     74 | 525000 | 525000 | 
| 54476593 |     74 | 525000 | 525000 | 
| 54476663 |     74 | 525000 | 525000 | 
| 54476833 |     74 | 525000 | 525000 | 
| 54476861 |     74 | 525000 | 525000 | 
| 54680371 |     74 | 525000 | 525000 | 
| 54698719 |     74 | 525000 | 525000 | 
| 54698759 |     74 | 525000 | 525000 | 
| 54847097 |     74 | 525000 | 525000 | 
| 55072681 |     74 | 530000 | 530000 | 
| 48091517 |     72 | 555000 | 555000 | 
| 48099379 |     72 | 555000 | 555000 | 
| 49232621 |     72 | 560000 | 560000 | 
| 49235027 |     72 | 560000 | 560000 | 
| 49304473 |     72 | 560000 | 560000 | 
| 45018373 |     72 | 565000 | 565000 | 
| 45434309 |     71 | 570000 | 570000 | 
| 46218209 |     72 | 575000 | 575000 | 
| 46262267 |     72 | 575000 | 575000 | 
| 46665911 |     72 | 580000 | 580000 | 
| 46780159 |     71 | 580000 | 580000 | 
| 54440039 |     73 | 580000 | 580000 | 
| 54440083 |     73 | 580000 | 580000 | 
| 54440291 |     73 | 580000 | 580000 | 
| 54440369 |     73 | 580000 | 580000 | 
| 54440389 |     73 | 580000 | 580000 | 
| 54442631 |     73 | 580000 | 580000 | 
| 54442643 |     73 | 580000 | 580000 | 
| 54444287 |     73 | 580000 | 580000 | 
| 54444323 |     73 | 580000 | 580000 | 
| 54444373 |     73 | 580000 | 580000 | 
| 54475163 |     73 | 580000 | 580000 | 
| 54627889 |     73 | 580000 | 580000 | 
| 54628121 |     73 | 580000 | 580000 | 
| 54629101 |     73 | 580000 | 580000 | 
| 50249491 |     72 | 585000 | 585000 | 
| 50377819 |     72 | 585000 | 585000 | 
| 54758579 |     73 | 585000 | 585000 | 
| 50575991 |     72 | 590000 | 590000 | 
| 54846683 |     73 | 595000 | 595000 | 
+----------+--------+--------+--------+

So, could you (and/or other experts) provide some guidance on how one might most optimally do this?

garo · 2012-01-08, 22:46

I think James has some P-1 credit computing code which can be useful. I'd say that redo P-1 when the marginal utility is the same as doing P-1 on a fresh exponent which is essentially never. And my opinion will not change until P-1 stops being short of resources.

axn · 2012-01-09, 03:52

Quote:

Originally Posted by garo

I think James has some P-1 credit computing code which can be useful. I'd say that redo P-1 when the marginal utility is the same as doing P-1 on a fresh exponent which is essentially never. And my opinion will not change until P-1 stops being short of resources.

P-1 will stop being short of resources when P95 default of 8MB is changed into something more meaningful (say 300MB or so). That 8MB figure is just way out of date.

chalsall · 2012-01-09, 05:16

Quote:

Originally Posted by garo

I think James has some P-1 credit computing code which can be useful. I'd say that redo P-1 when the marginal utility is the same as doing P-1 on a fresh exponent which is essentially never. And my opinion will not change until P-1 stops being short of resources.

I tend to agree with you.

chalsall · 2012-01-09, 05:19

Quote:

Originally Posted by axn

P-1 will stop being short of resources when P95 default of 8MB is changed into something more meaningful (say 300MB or so). That 8MB figure is just way out of date.

Agree.

But I do find it interesting that many of the candidates Spidy has seen have not had any P-1 done; even those which have had one (recent) LL test done.

This suggests to me that the "P-1 Hunters" are a valuable resource.

2012-01-08, 22:46	#30
garo Aug 2002 Termonfeckin, IE 5314₈ Posts	I think James has some P-1 credit computing code which can be useful. I'd say that redo P-1 when the marginal utility is the same as doing P-1 on a fresh exponent which is essentially never. And my opinion will not change until P-1 stops being short of resources.