Possible extention to the "GPU to 72 Tool" project?

[Dubslow]

Uhuh.



Anyways, chalsall,

Originally Posted by chalsall

Actually, that's not really an issue at all.

The system could simply keep a "store" of a certain number of exponents (say 100) which would have ordinarily been returned to PrimeNet.

As those who participate in the exercise (and it's still a proposal -- I won't implement unless it's agreed it won't have any harm and that it will be used) draw from the pool and take over ownership of the exponents, the system would simply top it up again (read: not release as many exponents back to PrimeNet as it would have ordinarily done).

The Perl and SQL to implement this would actually be less than the English describing it....

I'd still be surprised if you even managed one hit per day claiming an LL test. There's at best ten of us, I have almost the best proc of the line, and I can manage at most 4 per month. That's 40 per month total with some rather generous estimates. Hardly worth caching more than 20 at a time.

[lycorn]

We just are discussing quantities, not methods, it seems. Chalsall is right in that there is no issue. In order to "compress" the frontwave, the system would keep the lowest exponents (in a number tbd, and subject to adjustments depending on the number of participants in the sub-project) and release them to the trusted members.
Also note that people not participating as TFers in the GPU-to-72 project might qualify to receive these low exponents. For example, I don´t have a GPU useful to the project, but I have a fast processor and am willing to help out. I could register in GPU-to-72 but only to get low exponents to LL test.
What you people think of this possibility? We might get several people interested in having the lowest exponents to LL, even if they are not keen to Trial Factoring.

[davieddy]

Originally Posted by Dubslow

Uhuh.

See "useless posts" thread from #500 on to understand what
prompted this.
I can't understand why any mod (Tom?) could find
anything remotely suggestive about the admiral's daughter's
naval base being full of discharged seamen.

Pottymouth

fivemack: Whilst I may have gained an entirely deserved reputation as being swifter with the thread-move button and the great-big-hammer switch than some other mods, I haven't done anything on this thread. Perhaps I have been reading too much Hornblower, but I would have thought the average junior-ranking sailor would rather engage in carnal relations with a polar bear than with the admiral's daughter, the potential for injury and career-threatening consequence being less and the polar bear probably more cuddly.

[Dubslow]

Originally Posted by lycorn

We just are discussing quantities, not methods, it seems. Chalsall is right in that there is no issue. In order to "compress" the frontwave, the system would keep the lowest exponents (in a number tbd, and subject to adjustments depending on the number of participants in the sub-project) and release them to the trusted members.
Also note that people not participating as TFers in the GPU-to-72 project might qualify to receive these low exponents. For example, I don´t have a GPU useful to the project, but I have a fast processor and am willing to help out. I could register in GPU-to-72 but only to get low exponents to LL test.
What you people think of this possibility? We might get several people interested in having the lowest exponents to LL, even if they are not keen to Trial Factoring.

You make a fair point. Methodically, I don't think anybody sees a problem, and you do make the excellent point that this isn't limited to GPU'ers. Maybe it is worth a shot then, if we could get 15 people like you and me. I don't think that will happen; on the other hand, if chalsall is willing to implement this, then as long as we're/he's careful about not keeping assignments too long, this can only be a good thing (even if it's very little of a good thing).

If he does decide to implement it, then I would also request that we only get assigned LL's without P-1 done on them. Anybody with enough memory should be doing that.

One way to organize it might be to add three 'flags' (or whatever the appropriate term is) to each account, marking which work types they can do: GPU-TF, P-1, or LL. To get low P-1, check only that. To get LL, check only that. To get LL's that haven't been P-1'ed, check both checkboxes. (GPU-TF wouldn't affect those assignments.) Of course, I'm not a programmer, so I'm sure this won't work as is.

[chalsall]

Originally Posted by lycorn

Also note that people not participating as TFers in the GPU-to-72 project might qualify to receive these low exponents. For example, I don´t have a GPU useful to the project, but I have a fast processor and am willing to help out. I could register in GPU-to-72 but only to get low exponents to LL test.

Thanks lycorn. I thought I had explained (or, at least, inferred) that the LL Workers would not have to be GPU Workers (in fact, I assumed for the most part they wouldn't be), but obviously I hadn't clearly enough.

To be explicit, what I am proposing is simply to allow trusted, high-producing LL Workers to leverage on the work GPU to 72 is already doing. They benefit by getting the lowest available exponents to test, all TFed to a higher bit-level than normal (and soon, P-1ed properly). GIMPS benefits by having a greater number of lower exponents assigned to those who we know will complete the work in a reasonable time.

And, for those LL Workers who trust me enough, this could be fully automated for them by way of using gpu.mersenne.info as a Proxy for PrimeNet; the Workers would still receive all the credit.

All this code already exists -- I've been using it for years for my own Cluster controller.

[chalsall]

Originally Posted by Dubslow

You make a fair point. Methodically, I don't think anybody sees a problem, and you do make the excellent point that this isn't limited to GPU'ers. Maybe it is worth a shot then, if we could get 15 people like you and me. I don't think that will happen; on the other hand, if chalsall is willing to implement this, then as long as we're/he's careful about not keeping assignments too long, this can only be a good thing (even if it's very little of a good thing).

The number of assignments reserved would be implemented as a simple heuristic based on the number of requests actually made over time.

Originally Posted by Dubslow

If he does decide to implement it, then I would also request that we only get assigned LL's without P-1 done on them. Anybody with enough memory should be doing that.

I'm assuming you actually mean "with P-1 done on them".

I would counter instead each LL Worker is allowed to set preferences. Something like (in English) "Give me 10 LL tests to do below 50M TFed to at least 72 and P-1 done". If not that many are currently available with those characteristics, then the Worker could change their parameters such as not requiring P-1 done, lowering their TF limit, or raising their Exponent level.

Originally Posted by Dubslow

One way to organize it might be to add three 'flags' (or whatever the appropriate term is) to each account, marking which work types they can do: GPU-TF, P-1, or LL. To get low P-1, check only that. To get LL, check only that. To get LL's that haven't been P-1'ed, check both checkboxes. (GPU-TF wouldn't affect those assignments.) Of course, I'm not a programmer, so I'm sure this won't work as is.

Please don't take this the wrong way, but please leave the implementation issues to me. I actually sometimes dream in Perl....

[lycorn]

Originally Posted by chalsall

I'm assuming you actually mean "with P-1 done on them".

Hmmm... or not. A trusted LL tester is very likely to be a trusted P-1 tester, in that he is aware that a fair amount of memory shall be assigned to the P-1 test. We would be killing two birds with one stone, you see? Assigning exponents to people that would not only perform the LL test in a reasonable amount of time, but also previously perform a P-1 with an appropriate amount of memory.
I for one shall definitely choose that option.

[petrw1]

I assume anything you implement will be cleared with George (and/or Scott) first?

[Dubslow]

Originally Posted by lycorn

Hmmm... or not. A trusted LL tester is very likely to be a trusted P-1 tester, in that he is aware that a fair amount of memory shall be assigned to the P-1 test. We would be killing two birds with one stone, you see? Assigning exponents to people that would not only perform the LL test in a reasonable amount of time, but also previously perform a P-1 with an appropriate amount of memory.
I for one shall definitely choose that option.

Yes, I meant without. I can provide equal or better P-1 than everyone out there, so I would only want to take exponents without P-1 so that I can do it. It's my way of contributing to P-1 without giving up LL. (And yes, please implement as you like.)

As for confirmation with George/Scott, for one thing they're not around here very often, and for another we didn't go through them for the first stage of GPU to 72. If they complain we can always stop, and we would if they asked (right?). In the end any anonymous reservations made by the spider are either unreserved or claimed by on the of the workers, so none of these will be untracked or break PrimeNet in any way.

[chalsall]

Originally Posted by Dubslow

Yes, I meant without. I can provide equal or better P-1 than everyone out there, so I would only want to take exponents without P-1 so that I can do it. It's my way of contributing to P-1 without giving up LL. (And yes, please implement as you like.)

OK. Some LL Workers might want to take exponents which only HAD P-1 done because they don't have enough memory to do P-1 well. Simple to make the both options (plus "either way") available to requesters.

Originally Posted by Dubslow

As for confirmation with George/Scott, for one thing they're not around here very often, and for another we didn't go through them for the first stage of GPU to 72. If they complain we can always stop, and we would if they asked (right?). In the end any anonymous reservations made by the spider are either unreserved or claimed by on the of the workers, so none of these will be untracked or break PrimeNet in any way.

As P-1 said on the thread where the "Tool" was first discussed, George knows about this activity and hasn't complained.

And, absolutely -- George is God. If he says "stop", we stop. If he says "change your heuristics", we change.

I have a conference for the next three days. If I don't hear from George saying "Don't do that." before Friday, I'll begin implementing over the weekend.

[petrw1]

Originally Posted by Dubslow

I would only want to take exponents without P-1 so that I can do it.

If that (doing P-1 only) is an option I could help out there too.

I've got 9 cores available.