Reference material discussion thread

[kriesel]

Originally Posted by axn

v7 also has faster stage 2

That's not what I measured. See https://www.mersenneforum.org/showpo...9&postcount=29. For the same fft length / similar exponent, bounds, OS, system, GPU, etc., V7.2-53 stage 2 took significantly longer.
Thanks for using the discussion thread as intended BTW. (That is surprisingly rare among some other forumites.)

[drkirkby]

Originally Posted by SELROC

I'd say that you have to upgrade the bios as well.

Generally I would agree, bit I am aware of a possible reason why it might not be a good idea to update the BIOS for some people.

Some CPUs on eBay are either engineering samples of mainstream CPUs, or non-standard CPUs that were supplied to a specific customer - Amazon in particular often have non-standard CPUs. Apparently it is possible for a BIOS update to render a system unusable with a CPU that was previously working fine. So if one uses weird CPUs, and they are working, it might be better to not update the BIOS.

[drkirkby]

kriesel,
A comment on your post about worktodo.txt

https://www.mersenneforum.org/showpo...8&postcount=22

"tests_saved = integer number of future primality tests saved if a factor is found, usually 2 for a first test candidate, 1 for a double-check candidate, 0 if a sufficient P-1 has already been completed, or optionally up to 9 for aggressive P-1 factoring"

Although I have not verified it, from things I've seen written, this can be a real number, so is not restricted to integer values as stated. Someone wrote that mprime only takes two digits after the decimal point.

[kriesel]

Originally Posted by drkirkby

this can be a real number, so is not restricted to integer values as stated. Someone wrote that mprime only takes two digits after the decimal point.

Strictly speaking, mprime / prime95 would be supporting rational numbers x/10ⁿ. I've only ever seen single-digit integer values issued by PrimeNet or manual assignment page. I've run CUDAPm1 with several values.

[axn]

Originally Posted by kriesel

That's not what I measured. See https://www.mersenneforum.org/showpo...9&postcount=29. For the same exponent, bounds, OS, system, GPU, etc., V7.2-53 stage 2 took significantly longer.

Interesting. Do you have the screen output / logs from the v7 run (specifically with the maxAlloc enforced)?

I am interested in seeing all the internal parameters it chose.

[Uncwilly]

Originally Posted by kriesel

Strictly speaking, mprime / prime95 would be supporting rational numbers x/10ⁿ. I've only ever seen single-digit integer values issued by PrimeNet or manual assignment page.

Currently PrimeNet has been issuing only 1 and 2. But if the user sets the value to, say, 1.3 Prime95 will gladly calculate bounds based upon that value. I have taken to using values 1<x<1.4 when doing P-1 to free up exponents. I like finding factors.

[drkirkby]

Originally Posted by kriesel

Strictly speaking, mprime / prime95 would be supporting rational numbers x/10ⁿ. I've only ever seen single-digit integer values issued by PrimeNet or manual assignment page. I've run CUDAPm1 with several values.

I have only seen integers issued by the server. I convert the 2's to 1's. (Up until quite recently, I normally found that the P-1 factoring had been done by someone else on category 0 and 1 assignments, so tests_saved was 0. However, now I'm finding that's no longer the case, and it's usually necessary to do the P-1 factoring).

However, if one wants to use non-integer values for tests_saved, as Uncwilly chooses to use 1.4, tests_saved must be written as a decimal, and not as a fraction m/n. As a test to see if fractions were interpreted as fractions, I put these three lines in worktodo.add, then run mprime -c.

PRP=1,2,504575707,-1,76,14/10
PRP=1,2,504575749,-1,76,7/5
PRP=1,2,504575831,-1,76,1.4

After that, worktodo.txt contained the following lines.

PRP=CF4DE4E86F51E9B8ADD3DC07311E828F,1,2,504575707,-1,76,10
PRP=5DDE68758BD37D820D703388F5153630,1,2,504575749,-1,76,7
PRP=F7DD04FC451DCF6B7A9A2A38FDC828C1,1,2,504575831,-1,76,1.4

(All exponents have been unassigned, so the AIDs are no longer valid).

Anyway, I thought your entry about worktodo.txt could do with a slight tweak to reflect the fact that the numbers do not have to be integers.

[kriesel]

Originally Posted by drkirkby

if one wants to use non-integer values for tests_saved, as Uncwilly chooses to use 1.4, tests_saved must be written as a decimal, and not as a fraction m/n.

Anyway, I thought your entry about worktodo.txt could do with a slight tweak to reflect the fact that the numbers do not have to be integers.

My point regarding rationals not the larger set called reals is there is no way to enter irrational or transcendental numbers as input. I did not mean that ratios or any algebraic or function forms were accepted input. The source code shows %f format in sscanf. Any tests_saved value of finite precision put on a worktodo entry, entered by %f, or its approximate internal representation in binary, is a rational number. Some applications further restrict tests_saved to natural numbers or a small subset.

https://www.mersenneforum.org/showpo...8&postcount=22 has been modified repeatedly regarding tests_saved and with misc. edits.

[kriesel]

Originally Posted by axn

Interesting. Do you have the screen output / logs from the v7 run (specifically with the maxAlloc enforced)?

I am interested in seeing all the internal parameters it chose.

Log excerpts are added to the post linked earlier.

[axn]

Originally Posted by kriesel

Log exerpts are added to the post linked earlier.

I was looking for the stage 2 parameters used by v7. It's ok if you don't have it handy. Is preda aware of this regression?

[kriesel]

Originally Posted by axn

I was looking for the stage 2 parameters used by v7. It's ok if you don't have it handy. Is preda aware of this regression?

Added to https://www.mersenneforum.org/showpo...9&postcount=29. I can't say whether Preda has seen that, https://www.mersenneforum.org/showpo...35&postcount=2 updated in April, or
https://mersenneforum.org/showpost.p...&postcount=197 posted on March 7. The last gpuowl Github change was March 2.