[QUOTE=Christenson;261325]I had missed that the practical B1 on my machine is 660,000, and 660,000 Primorial is of, well, nontrivial, size.:blush:[/QUOTE]It's worse than that: for B1=660,000 P-1 actually computes a number whose exponent is much larger than 660,000 primorial.

The exponent to which it raises 3 is not just the primorial of B1, which is the product of all primes up to B1; it's the product of all [U]prime powers[/U] up to B1.

So, for B1 = 660,000, the exponent product includes 390625 = 5^8, instead of just 5.

[QUOTE=cheesehead;261367]It's worse than that: for B1=660,000 P-1 actually computes a number whose exponent is much larger than 660,000 primorial.

The exponent to which it raises 3 is not just the primorial of B1, which is the product of all primes up to B1; it's the product of all [U]prime powers[/U] up to B1.

So, for B1 = 660,000, the exponent product includes 390625 = 5^8, instead of just 5.[/QUOTE]

It wouldn't matter; we can directly represent the exponent, but not the result of the exponentiation, even if we build the one devices discussed in another thread to represent the list of all the primes up to 600 digits. Remember that the environmental impact statement for those outweighs the earth!

Was idly wondering if P-1 effort was keeping up with or falling behind the LL effort?

[QUOTE=Christenson;263589]Was idly wondering if P-1 effort was keeping up with or falling behind the LL effort?[/QUOTE]I'll venture that it will perpetually fall behind until P-1 gets the kind of computational breakthrough that GPUs have given TF over the last year.

The P-1 explanation page isn't terribly clear... is P-1 inherently seuqential, like LL tests, or can it be parallelized, like TF tests? Should I start a thread on the P-1 algorithm as homework?

Stage 2 can be done in parallel, but it requires the save file from stage 1.

How big is the save file? Would it be feasible to do stage 1 on a system with limited memory, then copy the save file to another user's system with enough memory to do stage 2?

Re doing p-1 on a GPU, would a GPU have enough memory for stage 2? If not you would still gain from doing stage 1 on a GPU and stage 2 on a CPU.

Chris K

[QUOTE=chris2be8;263617]How big is the save file?[/QUOTE]
[URL="http://www.mersenneforum.org/showpost.php?p=210670&postcount=36"]Here's an example of a save from GMP-ECM.[/URL] Size varies.
[QUOTE]Would it be feasible to do stage 1 on a system with limited memory, then copy the save file to another user's system with enough memory to do stage 2?[/QUOTE]Yes, that can be done. Prime95 saves to its own directory. I believe the documentation tells you the extensions of the different files. (Which files are P-1 etc.)
[QUOTE]Re doing p-1 on a GPU, would a GPU have enough memory for stage 2? If not you would still gain from doing stage 1 on a GPU and stage 2 on a CPU.[/QUOTE]I'm not sure about the current status of P-1 on GPUs.

Re: enough memory -- P95 requires 300M before (I think) before assigning P-1 work. A typical high-end GPU these days is going to have 768M or a Gig.

If stage 1 can't be sped up heavily by a GPU, but stage 2 can, then we will certainly want multiple systems feeding stage 1 save files to a single GPU.

I have to go penetrate P95 source code anyway. I note that even without the GPU speedup, I'm producing factors by P-1 for significantly less effort than two CPU-based LL tests.

[QUOTE=chris2be8;263617]How big is the save file?[/QUOTE]

5MB or more during stage 1. Double that once stage 1 is complete.

[QUOTE]Would it be feasible to do stage 1 on a system with limited memory, then copy the save file to another user's system with enough memory to do stage 2?[/QUOTE]

Very feasible; this is what I do:

Set the nighttime memory on the stage 1 machine to the amount you will be using on the stage 2 machine. You can't do this via the menu system if that's more memory than the stage 1 machine has, but you can edit local.txt directly. Leave the daytime memory at 8MB.

Put LowMemWhileRunning=prime95 (Windows) or LowMemWhileRunning=mprime (Linux) into your prime.txt file, to prevent the program for ever trying to use the high memory setting.

[QUOTE=lorgix;263609]Stage 2 can be done in parallel, but it requires the save file from stage 1.[/QUOTE]
Not quite right.
What is actually paralelized is the FFT algorithm. Both Stage 1 and 2 use FFTs to perform the calculations, so both Stages might benefit from the FFT paralelization (multi-threading). What can´t be paralelized, IIRC, are the GCD computations performed at the end of each Stage.