Double-Check LL (deprecated)

kdelisle2005 · 2023-02-10, 17:17

Hello.

see attached .jpg for manual assignment selection choices.

I have been doing only Double Check LL assignments since 2005. I prefer these because the run duration on my hardware was around 21 days. I also do manual assignment and that is where my question originates - what do I select for LL double-checks since it appears that double-check LL assignments are discouraged?

Should I be requesting assignments from "Double-check using PRP with proof" or "Double-check PRP tests"? I realize I can simply continue asis in the past and select "Double-check LL tests" but it seems there is a penalty for doing this. I was puzzled when I submitted my manual results that the reply said something about only receiving 35% of the GHz Days earned. Why is that happening?

I do not want to start a new set of result statistics.

thank you
kevin

moebius · 2023-02-10, 18:28

Double-check using PRP with proof

kriesel · 2023-02-10, 19:25

"Double check using PRP with proof" is DC of an LL first run with PRP/GEC/proof subsequent run.
Choosing that will get you some low much needed exponent DC work. This will be successful even on unreliable hardware. It will also typically generate a proof file allowing its confirmation of completeness and correctness at low cost.
PRP with proof generation is so much more reliable in practice that if you have the storage space to run a reasonable proof power, use that even to DC a previous first LL test.

"Double Check PRP tests" is DC of a PRP first run with another PRP run. PRP type and seed must match. Preferably the second PRP will generate a proof file that gets uploaded. Confirmation and reliability advantages like for the preceding choice.

"Double Check LL tests" is DC of an LL previous run with another LL run. If the seed values are the same and the users are different and the final res64 values match we hope that means both LL runs were correct. If they don't match at least one was wrong and a tiebreaker or more than one more run will be needed. There is no proof of correctness like with PRP/proof/Cert. There are some errors that produce random final res64s. There are also some errors that produce systematic wrong res64, which can match and be wrong. So this is the least reliable form of DC. It's also the lowest space and communication requirement, which can make it attractive for applications such as tiny-computer. Empirically determined error rate for LL testing ~2% per test at moderate exponent, 20% at 100Mdigit, apparently proportional to run time so essentially 100% at 1G-1Gdigit without special measures such as parallel runs with frequent manual cross checking of interim results.

Prime95 · 2023-02-10, 20:05

There is plenty of LL double-check work to be done (and most welcome).

I'm unaware of a 35% credit for LL DC. Asking James for his input.

kdelisle2005 · 2023-02-10, 20:58

Thank you for the clarification of the three double-check choices. Perhaps I will take one of each then compare the execution behaviors on a single PC (btw Ryzen 5 1600).

I should have a manual result to enter within the next 5-7 days and I will post the message that I receive about getting 1/3 of the GHz-days credit.

kevin

kriesel · 2023-02-10, 21:05

Kevin, do you have a capable GPU? There's lot of fun can be had there too.

Andrew Usher · 2023-02-11, 00:03

They should all run at practically the same speed for the same exponent. If you've been doing LL DCs for some time your computer is presumably giving reliable results on it, and you may continue. PRP/proof has only a small theoretical advantage and that is only realised if proofs are generated and uploaded every time - without a proof it is worth nothing as LL and PRP are different tests and do not verify each other.

The word 'deprecated' is probably left over from when first-time LL tests were still allowed as a choice. It seems to just confuse people now - given that the great majority of DCs are done LL, it seems silly to call them 'deprecated' or give less credit (for manual assignments only? Isn't the credit system separate from the assignment system?).

slandrum · 2023-02-11, 00:35

I have most of my machines set to "do whatever is most appropriate", and most of the assignments they get are LL-DC. There has been no reduction in credit awarded for LL-DC assignments, they get the full credit. Since the server is determining what's appropriate, I'd have to say that LL-DC is definitely not deprecated.

For my machines that are currently set just to LL-DC assignments, there is also no reduction in credit awarded.

There may be a reduction in credit for LL FTC, I haven't looked into that, there's only a couple people that I see sometimes doing new LL FTC assignments. Those are pretty worthless now because it's likely that no-one will do LL-DC on them, they'll be redone as PRP. The error rate on LL FTC at the wavefront makes it preferable to do PRP with error checking and certification than to do LL-DC there.

Andrew Usher · 2023-02-11, 17:59

Then it seems the word 'deprecated' should simply be removed from that menu. While any evidence of reduced credit should be shown, I rather doubt it as I don't think the server assigns credit based on assignment (or lack of it).

Also, that is probably understating the reliability of LL tests today. Compared to the data that the usually quoted figures of LL reliability were based on, an LL today is likely to

- Be done on a computer of at least average reliability
- Run faster, reducing the chance of memory corruption
- Always use the Jacobi check

so, while I can't give exact figures, I don't think there need be an upper limit to LL DCs.

chalsall · 2023-02-11, 21:00

Quote:

Originally Posted by Andrew Usher

... so, while I can't give exact figures, I don't think there need be an upper limit to LL DCs.

I enjoy playing with you et al I actually find the noise interesting to observe.

Not quite white; not quite pink. I must run a deep FFT on that heuristic.

There must be something there... Some effort expended...

1997rj7 · 2023-02-14, 04:24

If you get an LL double check through the manual assignment page, you get 1/3 credit for it.

If you get an LL double check from Primenet automatically, you get full credit.

Why?

2023-02-10, 17:17	#1
kdelisle2005 Mar 2006 Michigan 2⁴ Posts	Double-Check LL (deprecated) Hello. see attached .jpg for manual assignment selection choices. I have been doing only Double Check LL assignments since 2005. I prefer these because the run duration on my hardware was around 21 days. I also do manual assignment and that is where my question originates - what do I select for LL double-checks since it appears that double-check LL assignments are discouraged? Should I be requesting assignments from "Double-check using PRP with proof" or "Double-check PRP tests"? I realize I can simply continue asis in the past and select "Double-check LL tests" but it seems there is a penalty for doing this. I was puzzled when I submitted my manual results that the reply said something about only receiving 35% of the GHz Days earned. Why is that happening? I do not want to start a new set of result statistics. thank you kevin Attached Thumbnails

2023-02-10, 19:25	#3
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 7717₁₀ Posts	"Double check using PRP with proof" is DC of an LL first run with PRP/GEC/proof subsequent run. Choosing that will get you some low much needed exponent DC work. This will be successful even on unreliable hardware. It will also typically generate a proof file allowing its confirmation of completeness and correctness at low cost. PRP with proof generation is so much more reliable in practice that if you have the storage space to run a reasonable proof power, use that even to DC a previous first LL test. "Double Check PRP tests" is DC of a PRP first run with another PRP run. PRP type and seed must match. Preferably the second PRP will generate a proof file that gets uploaded. Confirmation and reliability advantages like for the preceding choice. "Double Check LL tests" is DC of an LL previous run with another LL run. If the seed values are the same and the users are different and the final res64 values match we hope that means both LL runs were correct. If they don't match at least one was wrong and a tiebreaker or more than one more run will be needed. There is no proof of correctness like with PRP/proof/Cert. There are some errors that produce random final res64s. There are also some errors that produce systematic wrong res64, which can match and be wrong. So this is the least reliable form of DC. It's also the lowest space and communication requirement, which can make it attractive for applications such as tiny-computer. Empirically determined error rate for LL testing ~2% per test at moderate exponent, 20% at 100Mdigit, apparently proportional to run time so essentially 100% at 1G-1Gdigit without special measures such as parallel runs with frequent manual cross checking of interim results. Last fiddled with by kriesel on 2023-02-11 at 02:14

2023-02-11, 00:35	#8
slandrum Jan 2021 California 3²×61 Posts	I have most of my machines set to "do whatever is most appropriate", and most of the assignments they get are LL-DC. There has been no reduction in credit awarded for LL-DC assignments, they get the full credit. Since the server is determining what's appropriate, I'd have to say that LL-DC is definitely not deprecated. For my machines that are currently set just to LL-DC assignments, there is also no reduction in credit awarded. There may be a reduction in credit for LL FTC, I haven't looked into that, there's only a couple people that I see sometimes doing new LL FTC assignments. Those are pretty worthless now because it's likely that no-one will do LL-DC on them, they'll be redone as PRP. The error rate on LL FTC at the wavefront makes it preferable to do PRP with error checking and certification than to do LL-DC there. Last fiddled with by slandrum on 2023-02-11 at 00:36 Reason: changed a word for clarity in meaning

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Double-Check	wombatman	Conjectures 'R Us	3	2016-08-29 20:46
Double Check	Unregistered	Information & Answers	3	2011-10-01 04:38
First check and double check llrnet servers.	opyrt	Prime Sierpinski Project	3	2009-01-02 01:50
Double-check check?	M0CZY	Software	15	2008-10-30 14:20
Double Check P-1	PhilF	PrimeNet	6	2005-07-03 14:36

2023-02-10, 18:28	#2
moebius Jul 2009 Germany 3²×7×11 Posts	Double-check using PRP with proof

2023-02-10, 20:05	#4
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 20076₈ Posts	There is plenty of LL double-check work to be done (and most welcome). I'm unaware of a 35% credit for LL DC. Asking James for his input.

2023-02-10, 20:58	#5
kdelisle2005 Mar 2006 Michigan 10000₂ Posts	Thank you for the clarification of the three double-check choices. Perhaps I will take one of each then compare the execution behaviors on a single PC (btw Ryzen 5 1600). I should have a manual result to enter within the next 5-7 days and I will post the message that I receive about getting 1/3 of the GHz-days credit. kevin

2023-02-10, 21:05	#6
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 7717₁₀ Posts	Kevin, do you have a capable GPU? There's lot of fun can be had there too.

2023-02-11, 00:03	#7
Andrew Usher Dec 2022 443 Posts	They should all run at practically the same speed for the same exponent. If you've been doing LL DCs for some time your computer is presumably giving reliable results on it, and you may continue. PRP/proof has only a small theoretical advantage and that is only realised if proofs are generated and uploaded every time - without a proof it is worth nothing as LL and PRP are different tests and do not verify each other. The word 'deprecated' is probably left over from when first-time LL tests were still allowed as a choice. It seems to just confuse people now - given that the great majority of DCs are done LL, it seems silly to call them 'deprecated' or give less credit (for manual assignments only? Isn't the credit system separate from the assignment system?).

2023-02-11, 17:59	#9
Andrew Usher Dec 2022 1BB₁₆ Posts	Then it seems the word 'deprecated' should simply be removed from that menu. While any evidence of reduced credit should be shown, I rather doubt it as I don't think the server assigns credit based on assignment (or lack of it). Also, that is probably understating the reliability of LL tests today. Compared to the data that the usually quoted figures of LL reliability were based on, an LL today is likely to - Be done on a computer of at least average reliability - Run faster, reducing the chance of memory corruption - Always use the Jacobi check so, while I can't give exact figures, I don't think there need be an upper limit to LL DCs.

2023-02-14, 04:24	#11
1997rj7 Sep 2003 2×179 Posts	If you get an LL double check through the manual assignment page, you get 1/3 credit for it. If you get an LL double check from Primenet automatically, you get full credit. Why?