P-1 factoring anyone?

[garo]

I think one of the main reasons TF is overpowered is because the Primenet/Prime95 P4 equivalent CPU speed system consistently underestimates the power of a computer. My Corei7 that runs at 3800MHz and has four cores shows up as 9.5 GHz P4 equivalent which is just plain wrong (unless that is the equivalent speed of one core in which case it would be in the ballpark). It runs 24/7 and has the number of hours option set to 24 as well. Looking at my other CPUs it seems that all processors that are Core2 or higher are treated as equivalent to a P4 at the beginning and their real speed is only reflected through the rolling average which increases over time.

[petrw1]

Originally Posted by garo

My Corei7 that runs at 3800MHz and has four cores shows up as 9.5 GHz P4 equivalent which is just plain wrong (unless that is the equivalent speed of one core in which case it would be in the ballpark).

I see similar results and have asked George about this before. I was told the Equivalent number is for each core. So it is suggesting your one core runs 9.5 times faster than the Ghz Day benchmark which would imply that each core is producing 9.5 points per day. However I am NOT seeing that ... but about half the number reported for my Duos / Quads. I have a Duo with an Equiv Factor of 4.2 for which each core produces just over 2 Ghz Day per Day; and a Quad with a Equiv Factor of 5 for which each core produces about 2.5 Ghz Day per Day. Hmmm, wait a minute. The Spec is based on a "theoretical 1 Ghz Duo" so maybe it is estimating for 2 cores. However, even my Single Core PCs Ghz Day Equivalent Factor are about double the points per day they produce. So maybe the interpretation is that both cores of this theoretical Duo combined produce 1 Ghz Day per Day - one half for each core. Then if I have a single core P4 producing 1.25 ponts per day then that is 2.5 times a core of the Theoretical PC. And a Quad cores that producing 2.5 ponts per day should have an Equiv factor of 5 (which it does). Does this "quess" work for you? Is each core producing 9.5 / 2 or 4.75 Ghz days?

Originally Posted by garo

Looking at my other CPUs it seems that all processors that are Core2 or higher are treated as equivalent to a P4 at the beginning and their real speed is only reflected through the rolling average which increases over time.

Yes I see that too; and especially with my new faster i5 it seems to take a long time. For example it started out estimating it would take 17 days for a DC when it was taking just over 5. Two weeks later the estimate has been adjusted to 11 days - still more than double.

[petrw1]

Originally Posted by petrw1

I see similar results and have asked George about this before. I was told the Equivalent number is for each core.

Or maybe I should have just read this post .....

http://www.mersenneforum.org/showpos...07&postcount=3

[markr]

A lot of people with a Core2 quad set it to run three cores LL/DC and one core TF. No idea how much effect this would have on PrimeNet's throughput.

[Uncwilly]

Originally Posted by markr

A lot of people with a Core2 quad set it to run three cores LL/DC and one core TF. No idea how much effect this would have on PrimeNet's throughput.

If they would do DC, LL, TF, and P-1 on the various respectve cores, that would be great.

[Brian-E]

Originally Posted by cheesehead

I meant to suggest that folks who currently specify "whatever makes sense" would change their preferences to specific types such as LL, DC, P-1 ... (the latter is what I meant by "that which I want to do") ...

... which is just what you seem to have done.

Yes, I understood what you wrote. The point I unsuccessfully tried to make clear was that while you were talking about people possibly changing to specific preferences if "whatever makes sense" were to be tweaked to give more P-1 and other under-resourced work-types, others - such as myself - have done that as a result of "whatever makes sense" not yet having been so tweaked - and we would go back to "whatever makes sense" if it were tweaked.

Surely many people who have "whatever makes sense" selected have done that because they are concerned that their precious cycles are being used optimally for the project.

The effect you warned about may well of course also occur with other contributors, but I wanted to mention the other side of the coin.

You know that PrimeNet's "whatever makes sense" doesn't mean simply "whichever category is most under-resourced from a forum participant's POV", don't you? IIRC it takes other factors into account, such as your CPU type and whatever-that-rating-of-how-PrimeNet-thinks-you've-done-so-far-is-called (not reliability, but something like that) in addition to George's fudge factors.

Yes, but the ultimate aim should still be to maximise productivity in the project as a whole. Of course a very important consideration is efficient use of the different machines that different people have and I understand that some machines are more efficient at particular work-types than others, but another consideration is that all work-types are sufficiently carried out so that new LL tests are not held up by shortage of factored exponents. It's that last consideration which - on the basis of what people are saying in this thread and others - is possibly not being adequately handled. And "do what makes sense" should by its own definition be addressing this.

When I've specified "whatever makes sense", I've never been given TF assignments, but maybe if I were still using my mid-1990s P75 to request assignments ...

Yes, my machine (Athlon 64-bit, only 3 years old, but even so just a single core), together with my limited hours per day (average 7 hours) of running, will be the reason why I was being given TF assignments when I had "do what makes sense" active. My first attempt to do something about this was to increase the memory I would allow it to use from 250 Mbyte to 400 Mbyte as a result of what people were saying about memory requirements for P-1. However this did not cause PrimeNet to start giving me P-1 assignments instead. Maybe 400 Mbyte is still insufficient for PrimeNet's liking - in which case tough luck because I cannot allow any more.
However, I don't believe I should have been getting TF instead of the much needed Double Checking which, according to those who will know, is still falling further and further behind. My Athlon does LL testing just as efficiently as TF if the rate of increase in my credit is to be believed.
Therefore I have now selected DC as my prefered option. (I had P-1 before, but that 400M limit might be harming things because it has been selecting slightly lower bounds than most other contributors factor to for similar size exponents, so I'll stop that possible harm.)

My motivation is entirely the efficient contribution to GIMPS. I would have thought that most people who select "do what makes sense" have that motivation. Therefore I urge the project administrators to tweak this option to help along those work-types which they believe need more attention.

[petrw1]

Originally Posted by Mr. P-1

..

... caught you ...

[petrw1]

101 leading edge P-1 completions in January ... 6 factors found.


About enough to keep all 4 cores of a good quad busy for 2 years.

[markr]

Here's one I hadn't seen before - I just submitted a factor from P-1 stage 1 using the manual results page, and it assumed it was found by TF. Previously I'd seen factors from P-1 on small exponents treated as though from ECM, and factors from P-1 stage 2 treated as though from stage 1.

Processing result: M51657311 has a factor: 465878761401323109391, AID: 3CDB06AEA55BB05AAE007096B4E093D1
Insufficient information for accurate CPU credit.
For stats purposes, assuming factor was found by trial factoring using prime95.
CPU credit is 0.3370 GHz-days.

TF would have found it - it was a 69 bit factor for a 51M exponent. The manual results page seems to only use the sizes of the exponent & factor, so it makes sense.

To keep my stats straight, maybe the way to go is to fool the client into making a prime.spl file & transfer it to a connected machine. Pity we can't upload a prime.spl file the same as a results.txt file.

[cheesehead]

Originally Posted by markr

Processing result: M51657311 has a factor: 465878761401323109391, AID: 3CDB06AEA55BB05AAE007096B4E093D1
Insufficient information for accurate CPU credit.
For stats purposes, assuming factor was found by trial factoring using prime95.
CPU credit is 0.3370 GHz-days.

TF would have found it - it was a 69 bit factor for a 51M exponent.

That's what will occasionally happen under the current scheme of having PrimeNet assign TF up to one bit level short of limit (69, for 51M range), then assigning P-1, then assigning TF for the last bit level. Thus, PrimeNet's assumption of TF was at odds with its assignment algorithm.

(That interrupted-TF scheme is more efficient in the long run, but will occasionally turn out to have been suboptimal in certain cases, as this shows.)

PrimeNet should be tweaked to use the same algorithm for assuming factoring method as it uses for making assignments in the exponent's range!

... until a more direct fix of having the client specify (and PrimeNet use for calculating credit) the actual factor-finding method in the factor-found report message, that is.

[lycorn]

Originally Posted by markr

Location: Looking for a new avatar

Will this one do?