error rates and P-1 test

[drakkar67]

With a speed boost from BOINC we almost passed 5000K for LLR tests. On the other hand doublecheck passed 1450K.
Can we learn the error rates and tests that need triple check regularly, ie. once a moth.
By the way do you think P-1 tests effective from this point?
Personally I want to make some trials on the reserved ranges for k=79309. Do you have any idea, what boundaries should I have to use for 5400K-7000K range? And which program is most effective for p-1 tests?

thanks
drakkar67

[hhh]

If you want to do some effective P-1, head over to seventeenorbust. For PSP, it will not be effective soon.

Why?

The program of choice to do P-1 is prime95, (or mprime, if Linux). As you might know, there are two boundaries, B1, and B2.

If you want to choose them yourself, You put the line

Pminus1=k,2,n,1,B1,B2,0

into the worktodo.ini

If you want to let them to be chosen by prime95, you put in

Pfactor=k,2,n,1,sievedepth,factorworth

2^sievedepth is how deep we sieved, i.e. currently about sievedepth=52, soon 53. Factorworth is how much LLRtests are saved by one factor, so something between 1 and 2, as primes might be found.
put in 2, for testing purposes, or "1.9".

So your line wouldlook like

Pfactor=90527,2,5049623,1,53,2,0

Don't forget to allocate memory (options/CPU) and let Prime95 start. It will say something like 90527*2^5049623+1 does not need factoring.

The reason is: We sieved that deep that P-1 isn't worth it.

For the seventeenorbust stuff, this isn't true, as the tests take longer. If you are interested, tell me, I figured out a way how to run stage1 on one Core and stage 2 on another, maximizing the efficiency, if you have plenty of memory.
Any help will be appreciated, as I feel a little lonely over there.

H.

[ltd]

Error rates in the region around 1.45M are a little bit tricky due to the fact the in that region there was a switch within the llrnet client which leads to not compareable residues. The problem with that is that you can not see from the result which client was used. Our tests with k=168451 show that the swithover was completed around 1.65M. From that point on the error rates are meaningfull again. In the region between 1.65M and 2.3M for that k we have an overall error rate of 5.2%.
But in there are 20 results from a user where there is a possibility that he still had used an old client. If we remove his results from the view we end up with an error rate of 3.6%. I will analyse the results further to see which eror rate is the correct one.

In the background there is already a dc llrnet server running but due to personal time restrictions i have not made an anouncement about it. So stay tuned for that.

For p-1 tests the easiest and as far as i know fastest way would be to us prime95. My tests with that software show that it is not usefull to run p-1 at the moment. I made some tests with n around 6.5M 3 months ago and at that n the software only starts a tests if i force the automatic B1/B2 calculation with a parameter value that says a factor should be worth 2.6
PRP tests. And now we have already sieved much deeper.
My guess is that with a reasonalbe factor value around 1.6 to 1.8 at the moment p-1 tests would be usefull for tests around 9M. (This is only a guess it could be even higher or a little bit lower)

Hope that helps.

Edit: while i was typing hhh already gave parts of my answer.

[Prime95]

Originally Posted by hhh

Factorworth is how much LLRtests are saved by one factor, so something between 1 and 2, as primes might be found.
put in 2, for testing purposes, or "1.9".

If PSP does no double-checking then put in 1 for the factorworth (if you find a factor then you will save exactly 1 LLR test). If PSP does do double-checking then you have to guess if a prime will be found for that k (one test saved) before the double-checking wave reaches this n (two tests saved). I would guess the double-checking wave is sufficiently behind that a factor worth of 1.3 would be more appropriate (30% chance that a factor found will save 2 LLR tests).

[hhh]

Originally Posted by Prime95

If PSP does no double-checking then put in 1 for the factorworth (if you find a factor then you will save exactly 1 LLR test). If PSP does do double-checking then you have to guess if a prime will be found for that k (one test saved) before the double-checking wave reaches this n (two tests saved). I would guess the double-checking wave is sufficiently behind that a factor worth of 1.3 would be more appropriate (30% chance that a factor found will save 2 LLR tests).

PSP is doing double checking, and you are right. But if you want to prove that it P-1 is not worth it, and you put in 2 and get the "not worth it"-message, you are on the safe side.

Besides, George, you told me once that the formula to determine B1 and B2 is extremely difficult, and I didn't insist. With the upcoming of dualcore processors, shouldn't be implemented some feature that *automatically* puts stage1 in one core, stave2 in the other, and gives plenty of memory to the second one? And mustn't the formula be adapted to such behaviour?
(I figured it would be possible to get such a behaviour via the "-Ax" swich, ).
H.

[Prime95]

Originally Posted by hhh

With the upcoming of dualcore processors, shouldn't be implemented some feature that *automatically* puts stage1 in one core, stave2 in the other, and gives plenty of memory to the second one? And mustn't the formula be adapted to such behaviour? .

Version 25 makes some progress along those lines. You only run one prime95 to keep both cores busy (no more -An). If both threads are doing P-1 then the first thread to get to stage 2 gets all the available memory. When the second thread wants to do stage 2 then instead it starts stage 1 on the next work unit in its queue. When the first thread completes stage 2, the second thread goes back and finishes off its first work unit.

Nice in theory, but there have been several complaints that it doesn't work well in all cases.

The B1/B2 formula doesn't need changing. Instead of running two prime95's with 512MB you run one with 1GB. The calculation uses the higher memory allowance to compute a slightly faster runtime (thus making P-1 effective a little sooner).

[drakkar67]

Thanks for the answers. Seeems we have more than a year to think about p-1 tests.
By the way is there a formula to slow down LLR a little, like the throttle function of p95? It's getting hotter here and we mostly use notebooks.

[hhh]

Originally Posted by drakkar67

Thanks for the answers. Seeems we have more than a year to think about p-1 tests.

For PSP, yes. For SoB, no; there will be PRP tests without P-1 over there.

Originally Posted by drakkar67

By the way is there a formula to slow down LLR a little, like the throttle function of p95? It's getting hotter here and we mostly use notebooks.

Dunno. Have you thought about underclocking or undervolting you notebook? There are a couple of tools out for that stuff, it works fine for my ThinkPad (don't know what kind of computer you have).

Cheers, H.

[drakkar67]

I don't know those tools. I have 4 C2D notebooks; one of them 12.1 inches which is the hottest.

[hhh]

Well, I could as well have given the hints right away:
Try googleling NHC or RMClock. I use the first one which is nicer to watch, but the latter might be more powerful. The best place to get started might be a userforum specific to your laptop brand.

My Core 2 Duo, when idle, is clocked at 1GHz, with 0.95V and 55°C, without fan. When on load, it goes up to 1.83GHz, 1.1V instead of 1.5V, and 67°C with fan level 3, which is still OK.

There might be issues and patches etc. you have to consider specific to processors, chipsets, and other specifications.

H.