[QUOTE=Prime95;170336]No. The primary source of did-not-bother-with-P-1-tests are manual testers. I cannot control what they type into worktodo.txt. The other source would be pre-v21(?) clients. All v24 and earlier clients look the same to the new server.[/QUOTE]

OK here are two suggestions:

1. Program the server to hand out P-1ed exponents (if there are any available) to clients requesting tests/doublechecks with insufficient memory to do stage 2, even if they are not the smallest available, and perhaps also to all Pre-V25 clients, and to hand out non-P-1ed exponents (if there are any, which would only be an issue for doublechecks) to clients which do have sufficient memory.

2. Program the server to hand out P-1ed exponents by default to manual testers, perhaps with an option to specifically request non-P-1ed exponents.

The obvious alternative is to not hand out un-P-1-ed exponents for LL tests at all...

[QUOTE=ckdo;170400]The obvious alternative is to not hand out un-P-1-ed exponents for LL tests at all...[/QUOTE]

That's the ideal. Unfortunately there aren't enough of them. That's why many of us recommend that those able to commit reasonable amounts of memory for a reasonable fraction of the time should choose P-1 factoring work to best help the project., and those not able to commit reasonable amounts of memory for any time at all should do doublechecks or TF.

[QUOTE=Mr. P-1;170409]That's why many of us recommend that those able to commit reasonable amounts of memory for a reasonable fraction of the time should choose P-1 factoring work to best help the project.[/QUOTE]

On my way... :spot:

1 Q9550 core starting beginning of next week.
1 PIV 3.4 starting early next week.
1 Q6600 core starting mid next week.

Possibly more to follow shortly ...

[quote=ckdo;170400]The obvious alternative is to not hand out un-P-1-ed exponents for LL tests at all...[/quote]Will you expand upon that, please? Do you mean that if there are no P-1ed exponents for LL tests, first-time-LL-requestors are to be informed that no exponents are available for first-time-LL-assignment, then to be given non-LL (but not P-1, either, presumably) assignments instead, or are they to be given DC assignments instead?

[QUOTE=petrw1;170419]On my way... :spot:

1 Q9550 core starting beginning of next week.
1 PIV 3.4 starting early next week.
1 Q6600 core starting mid next week.

Possibly more to follow shortly ...[/QUOTE]

That's great. On the other hand, you'll probably outstrip me on the P-1 rankings in very short order. That sucks.

Bear in mind that one core doing P-1 24/7 can feed many tens of comparable cores doing LL tests, so this is really well-worthwhile

Some tips:

Stage 2 takes longer than stage 1, how much longer depends upon the available memory. If you can ensure that nighttime memory is available for a sufficiently long period of time, you could set daytime memory to 8MB to ensure that stage2 is only done at night. OnlyRunStage2WithMaxMemory=1 in prime.txt option has the same effect.

For multi-core processors, you can use the MaxHighMemWorkers local.txt option to limit the number of simultaneous stage 2s. If you decide to allow some daytime stage 2, you could use the "value during list-of-times else value" format to allow more threads to do stage 2 at night than during the day.

Even if you have max memory available all the time, you might still use MaxHighMemWorkers with the "value during list-of-times else value" format to fine-tune your memory usage. For example on a quad core, allowing only 2 cores to do stage 2 at once would lead to an accumulation of uncompleted work, while allowing 3 will probably result in lengthy periods where just 1 core is doing stage 2. This is inefficient; you want it to doing stage 2 on at least 2 cores all the time. You might achieve this, by limiting it to two cores for four days (say) in the week and 3 for the remaining three days.

Be careful though, that you don't create a situation where uncompleted stage 2 work is allowed to accumulate indefinitely.

[QUOTE=Mr. P-1;170458]That's great. On the other hand, you'll probably outstrip me on the P-1 rankings in very short order. That sucks.

Bear in mind that one core doing P-1 24/7 can feed many tens of comparable cores doing LL tests, so this is really well-worthwhile

Some tips:

Stage 2 takes longer than stage 1, how much longer depends upon the available memory. If you can ensure that nighttime memory is available for a sufficiently long period of time, you could set daytime memory to 8MB to ensure that stage2 is only done at night. OnlyRunStage2WithMaxMemory=1 in prime.txt option has the same effect.

For multi-core processors, you can use the MaxHighMemWorkers local.txt option to limit the number of simultaneous stage 2s. If you decide to allow some daytime stage 2, you could use the "value during list-of-times else value" format to allow more threads to do stage 2 at night than during the day.

Even if you have max memory available all the time, you might still use MaxHighMemWorkers with the "value during list-of-times else value" format to fine-tune your memory usage. For example on a quad core, allowing only 2 cores to do stage 2 at once would lead to an accumulation of uncompleted work, while allowing 3 will probably result in lengthy periods where just 1 core is doing stage 2. This is inefficient; you want it to doing stage 2 on at least 2 cores all the time. You might achieve this, by limiting it to two cores for four days (say) in the week and 3 for the remaining three days.

Be careful though, that you don't create a situation where uncompleted stage 2 work is allowed to accumulate indefinitely.[/QUOTE]

Thanks for the tips but I don't intend to have more than 1 core per PC on P-1. If I decide to add more it will be 1 core on one or both Duos.

Wayne "I P-1 freely"

[quote=cheesehead;170429]Will you expand upon that, please?[/quote]

I was thinking along these lines:

- Hand out the lowest available exponent with no P-1 done when P-1 is requested (i.e. 30m instead of 50m at the moment)
- Hand out exponents with P-1 done when LL is requested (last bit of TF done or not) when LL is requested (there are around 9400 of these currently below 79.3m)
- When we run out of exponents with P-1 done, hand out P-1 assignments if the client can handle them (v25 & 200 meg+ or whatever) when LL is requested, and DC otherwise (or LL on exponents with no P-1 done, but I wouldn't take that risk, personally)

Additionally,

- educate users towards updating their clients and choosing reasonable memory settings and
- consider that "what makes most sense" might actually be P-1 if the client can handle it.

Just my 2 cents, of course.

P.S.: I have 3 cores doing P-1 24/7 currently.

[QUOTE=ckdo;170484]I was thinking along these lines:

- Hand out the lowest available exponent with no P-1 done when P-1 is requested (i.e. 30m instead of 50m at the moment)
- Hand out exponents with P-1 done when LL is requested (last bit of TF done or not) when LL is requested (there are around 9400 of these currently below 79.3m)
- When we run out of exponents with P-1 done, hand out P-1 assignments if the client can handle them (v25 & 200 meg+ or whatever) when LL is requested, and DC otherwise (or LL on exponents with no P-1 done, but I wouldn't take that risk, personally)[/quote]

This isn't a bad idea. However, instead of handing out P-1s to client requesting LLs, give them un-P-1ed LLs to do. I agree that clients not known to be P-1 stage 2 capable should only get P-1ed exponents, even if that means giving them doublechecks.

This would slow down the rate of testing and speed up doublechecking. But we're already well-advanced on testing.

Lol, I don't really get the flow of the conversation here but I'm really interested to understand what's happening.

I would like to ask what is the full process that an exponent has to go through. I'll just list down what I think happens and you guys can correct or add on to my list.


1. PrimeNet removes exponent with composite [I]n[/I] (where [I]n[/I] is from 2^[I]n[/I]-1).

2. Trial factoring done by PrimeNet up to 62 bits?

3. TF assigned to users.
63 bits?
64 bits?
...
Up to 67 bits?
Up to 68 bits??
Up to 69 bits???
I remember that last time, I had to perform TF from 63 bits all the way to 67 bits but PrimeNet v5 now assigns TF bit by bit.

4. P-1. I remember that last time, users who wanted to perform LL had to perform P-1 (both Stage 1 and 2) before they started LL. Everything was done by one user. But I gather from the previous posts that now, users can perform P-1 Stage 1 without having to perform Stage 2?

5. LL. What if the result is returned with an error reported by Prime95? It becomes a suspect LL and has to undergo LL again. But if the next LL has the same result as the first, does DC still need to be done?

6. DC. Check if the residue matches the residue of the LL. If not, reassign DC until there are two residues that match.

End.


Thanks.

1 Year update....
 

It was one year ago today on April 22, 2008 that I started this thread.
GIMPS was running V4.
I was still only dreaming about owning a Quad.
We only had 44 known Mersenne Primes
There was NO Financial Crisis.
[B][U]There were 9576 Exponents < 10M Digits yet to have 1 successful LL test.[/U][/B]

Now one year later:
GIMPS V5 is running smoothly.
I have my Q9550 for 5 months now.
2 more Mersenne Primes discovered.
Everyone is broker (to some degree)
[B][U]There are only 3387 exponents <10M digits remaining.[/U][/B]

That is 35.37% of the original number.

And the good news is there has been a recent concerted effort in this range; the count is dropping faster. Including a reduction of 483 in the last 20 days or 24 per day. A sustained rate of 24 per day gets it done before the one year anniversary of V5 in October. Only 14 per day gets it done before year end A-N-D determines (subject to a DC) whether the first 44 MPs are the only 44 MPs < 10M digits.

Go, go, go!!! :skiing: