mersenneforum.org Going back to PRP
 Register FAQ Search Today's Posts Mark Forums Read

 2021-03-08, 19:56 #1 JuanTutors     Mar 2004 22·33·5 Posts Going back to PRP I have two 100M digit assignments that, after starting the PRP tests, I've noticed they have not gotten P-1 done (and on some computers, I never get P-1 even when P-1 was never done on the exponents). The assignments are at 24.0% and 31.8%, respectively. Is it at all worth doing P-1 factoring on either of these assignments at this point?
2021-03-08, 20:21   #2
ewmayer
2ω=0

Sep 2002
República de California

19×613 Posts

Quote:
 Originally Posted by JuanTutors I have two 100M digit assignments that, after starting the PRP tests, I've noticed they have not gotten P-1 done (and on some computers, I never get P-1 even when P-1 was never done on the exponents). The assignments are at 24.0% and 31.8%, respectively. Is it at all worth doing P-1 factoring on either of these assignments at this point?
Which program are you using? If gpuowl, it's been doing a combined p-1 stage 1 and first-few-%-of-PRP since v7. That means that such a run starts out looking like a PRP, but the first few Miterations of the PRP will run 10-20% slower than usual because it's also doing some needed auxiliary computations to generate a p-1 stage residue. For a given stage 1 bitdepth B1, around 1.4*B1 iterations into the PRP test you'll see it then doing a GCD of the stage 1 residue to check for a factor. If it finds one, it reports it and moves to the next assignment, otherwise it proceeds with p-1 stage 2, and should that also fail to turn up a factor, it reports the p-1 failure and the bounds used to the server and completes the already-begun PRP. After completion of such a run, you should see credit for both p-1 and the PRP on the exponents status page.

If that's not what's happening in your case, can you give us an example for an exponent for which your run didn't do p-1, even though it should have?

2021-03-08, 20:24   #3
JuanTutors

Mar 2004

54010 Posts

Quote:
 Originally Posted by ewmayer Which program are you using? If gpuowl, it's been doing a combined p-1 stage 1 and first-few-%-of-PRP since v7. That means that such a run starts out looking like a PRP, but the first few Miterations of the PRP will run 10-20% slower than usual because it's also doing some needed auxiliary computations to generate a p-1 stage residue. For a given stage 1 bitdepth B1, around 1.4*B1 iterations into the PRP test you'll see it then doing a GCD of the stage 1 residue to check for a factor. If it finds one, it reports it and moves to the next assignment, otherwise it proceeds with p-1 stage 2, and should that also fail to turn up a factor, it reports the p-1 failure and the bounds used to the server and completes the already-begun PRP. After completion of such a run, you should see credit for both p-1 and the PRP on the exponents status page. If that's not what's happening in your case, can you give us an example for an exponent for which your run didn't do p-1, even though it should have?
I don't even know what gpuowl is ... I have been using Prime95 v30.3b6

(And my apologies for posting in the wrong section.)

Last fiddled with by JuanTutors on 2021-03-08 at 20:25

2021-03-08, 20:56   #4
JuanTutors

Mar 2004

22×33×5 Posts

Quote:
 Originally Posted by ewmayer Which program are you using? If gpuowl, it's been doing a combined p-1 stage 1 and first-few-%-of-PRP since v7. That means that such a run starts out looking like a PRP, but the first few Miterations of the PRP will run 10-20% slower than usual because it's also doing some needed auxiliary computations to generate a p-1 stage residue. For a given stage 1 bitdepth B1, around 1.4*B1 iterations into the PRP test you'll see it then doing a GCD of the stage 1 residue to check for a factor. If it finds one, it reports it and moves to the next assignment, otherwise it proceeds with p-1 stage 2, and should that also fail to turn up a factor, it reports the p-1 failure and the bounds used to the server and completes the already-begun PRP. After completion of such a run, you should see credit for both p-1 and the PRP on the exponents status page. If that's not what's happening in your case, can you give us an example for an exponent for which your run didn't do p-1, even though it should have?
Here is an example of an exponent that has not had any P-1 done. As I mentioned, I don't know what gpuowl is, so I've never used it. https://www.mersenne.org/report_expo...2292529&full=1

 2021-03-08, 21:43 #5 ewmayer ∂2ω=0     Sep 2002 República de California 2D7F16 Posts Ah, you're doing 100Mdigit work - I don't know what the default p-1 protocol for those is supposed to be. PMing George to see if he can clarify.
2021-03-08, 22:39   #6
Prime95
P90 years forever!

Aug 2002
Yeehaw, FL

19·397 Posts

Quote:
 Originally Posted by JuanTutors Here is an example of an exponent that has not had any P-1 done. As I mentioned, I don't know what gpuowl is, so I've never used it. https://www.mersenne.org/report_expo...2292529&full=1
If you got the assignment from the server and did not edit the worktodo.txt, then prime95 should know that the exponent has not had P-1 done. Prime95 will then use your current memory settings to decide if P-1 will be useful -- that is, your chance of finding a factor is worth investing the CPU time.

I suppose if you're not letting prime95 use enough memory to do a useful P-1 stage 2 (prime95's default setting), then prime95 could well have decided not to run P-1 given the deep trial factoring already done.

2021-03-08, 22:41   #7
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

5,437 Posts

Quote:
 Originally Posted by JuanTutors I have two 100M digit assignments that, after starting the PRP tests, I've noticed they have not gotten P-1 done (and on some computers, I never get P-1 even when P-1 was never done on the exponents). The assignments are at 24.0% and 31.8%, respectively. Is it at all worth doing P-1 factoring on either of these assignments at this point?
Yes.
Set allowed memory usage by P-1/ECM to something reasonable. More is often better, up to a point. Default prime95 settings are generally too low for adequate P-1 bounds. They're set conservatively low to create less trouble on low-ram systems IIRC.
See https://www.mersenne.ca/exponent/332292529 for suitable bounds.
I often give it 1/2, 3/4 or more of system ram. The attached example is a 16GB system which also runs multitab web browser & Colab instances, lots of remote desktop sessions, office apps, etc.
See also https://www.mersenneforum.org/showpo...8&postcount=22, https://www.mersenneforum.org/showpo...4&postcount=17, and (even though it's gpu-oriented, its conclusions also apply in cpu P-1) https://www.mersenneforum.org/showpo...9&postcount=20

Seems to me this thread belongs in the software subforum, or a prime95 subforum, not the PrimeNet subforum.
Attached Thumbnails

Last fiddled with by kriesel on 2021-03-08 at 23:23

2021-03-09, 02:50   #8
Prime95
P90 years forever!

Aug 2002
Yeehaw, FL

19×397 Posts

Quote:
 Originally Posted by kriesel Upgrade to v30.4b9.
30.4b9 has an issue for PRP testing (but not until you reach the last 50 iterations).

2021-03-09, 04:18   #9
JuanTutors

Mar 2004

22×33×5 Posts

Quote:
 Originally Posted by Prime95 30.4b9 has an issue for PRP testing (but not until you reach the last 50 iterations). 30.5 should be ready soon.
Thanks! I'll look out for it and for any bug fixes. Now to the question of whether it's worth doing P-1 on an exponent that is 31.8% done with PRP, is it worth it? Could/should I set the number of tests saved to a decimal like 1 - 0.318 = .612?

2021-03-09, 06:25   #10
kriesel

"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

124758 Posts

Quote:
 Originally Posted by Prime95 30.4b9 has an issue for PRP testing (but not until you reach the last 50 iterations). 30.5 should be ready soon.
Yikes! I've already migrated my fleet. Based on posts indicating v30.4 stable, a month ago, and a specific recommendation to migrate P-1 activity, which occupies a worker on most systems. Most of my systems run mixed work types among workers. Can work in progress safely be rolled back to v30.3b6?
The issue identified affects PRP only, not PRP-CF or LL or P-1?

"I consider this pretty stable software." https://mersenneforum.org/showpost.p...51&postcount=1
"30.4 seems stable enough to make the transition." https://mersenneforum.org/showpost.p...7&postcount=88

Quote:
 Originally Posted by JuanTutors Thanks! I'll look out for it and for any bug fixes. Now to the question of whether it's worth doing P-1 on an exponent that is 31.8% done with PRP, is it worth it? Could/should I set the number of tests saved to a decimal like 1 - 0.318 = .612?
From undoc.txt, seems to indicate reals are supported for num_primality_tests_saved in prime95.
Code:
You can do "optimal" P-1 factoring of k*b^n+c by adding lines to worktodo.txt:
Pfactor=k,b,n,c,how_far_factored,num_primality_tests_saved
For example, Pfactor=1,2,10000157,-1,64,2.0
I think that is not the case in some other GIMPS programs.

Last fiddled with by kriesel on 2021-03-09 at 06:55

2021-03-09, 09:55   #11
Prime95
P90 years forever!

Aug 2002
Yeehaw, FL

19·397 Posts

Quote:
 Originally Posted by kriesel Yikes! I've already migrated my fleet. Based on posts indicating v30.4 stable, a month ago, and a specific recommendation to migrate P-1 activity, which occupies a worker on most systems..

Stick with 30.4 but keep an eye on any boxes nearing PRP completion. No wrong answers are produced, just repeated roundoff errors. This makes it unsuitable for set-and-forget installation.

This is an example of why I wait a fairly long while before making a seemingly stable release the version available on the mersenne.org download page.

Last fiddled with by Prime95 on 2021-03-09 at 09:57

 Similar Threads Thread Thread Starter Forum Replies Last Post jwaltos Lounge 17 2020-03-13 08:11 tServo Lounge 3 2015-12-12 05:36 Nipal Information & Answers 3 2013-09-04 18:20 RickC PrimeNet 12 2011-08-14 02:33 mdettweiler No Prime Left Behind 0 2009-12-27 15:10

All times are UTC. The time now is 02:41.

Fri Aug 6 02:41:57 UTC 2021 up 13 days, 21:10, 1 user, load averages: 2.44, 2.40, 2.32