P-1 factoring anyone?

[db597]

Originally Posted by petrw1

When I reduced my P-1 RAM from 2500 to 1000 the chances of finding a factor reduced from 7.66% to 7.55%.

I suppose it's the law of diminishing returns. Difference between 200MB and 400MB going to be different from 1000MB and 2500MB.

[James Heinrich]

There are two things at play when changing the amount of RAM available for P-1:
a) for a given set of B1/B2 bounds, more RAM will allow the process to run faster
b) because of (a), if the amount of available of RAM is increased, Prime95 will select larger bounds to increase the chance of finding a factor, while maintaining the amount of time spent on P-1.

As an example (with made-up numbers), if you run P-1 on M123456 with 100MB allocated, Prime95 will select B1/B2 that will give you a 5.00% chance of finding a factor, and take 24 hours to complete. If you increase the amount of RAM from 100MB to 500MB, Prime95 could complete the P-1 in 22 hours with the same bounds, but it will usually prefer to adjust the bounds so it still takes about 24 hours to run, but you now have a 5.50% chance of finding a factor.

The bounds will be set based on the maximum of daytime/nighttime available memory. If the memory allocation is changed during stage1, Prime95 will proceed with the newly-chosen bounds. If the memory allocation is changed after stage2 has started, however, Prime95 will make use of the new amount of memory but not re-choose the bounds it uses.

[James Heinrich]

Originally Posted by James Heinrich

...much more than 500MB isn't necessary for now (but will be in the future years as exponents get larger).

As an illustration, P-1 on 100,000,000-digit numbers (e.g. M332203901 that I'm working on) takes 838MB plus 163MB for each relative prime it processes in each batch). So, to process 10 relative primes you need >=2468MB available.

Fortunately, only crazy people like myself are working on P-1 this far up the available-work spectrum right now; by the time it's mainstream these hardware requirements will seem trivial

[S00113]

Originally Posted by James Heinrich

As an illustration, P-1 on 100,000,000-digit numbers (e.g. M332203901 that I'm working on) takes 838MB plus 163MB for each relative prime it processes in each batch). So, to process 10 relative primes you need >=2468MB available.

Fortunately, only crazy people like myself are working on P-1 this far up the available-work spectrum right now; by the time it's mainstream these hardware requirements will seem trivial

The PrimeNet server must consider me crazy!

Server assigned P-1 factoring work.
Got assignment 5F100DBBA223AD8301246A47D1826BA6: P-1 M332194021
Sending expected completion date for M332194021: Feb 23 2009

The machine has been working a while, and mprime uses 3480 MiB at night, 2048 MiB daytime.

[James Heinrich]

Originally Posted by S00113

The PrimeNet server must consider me crazy!

Did PrimeNet actually assign you that number with P-1 as your preferred worktype, or did you manually add that? I would find it unexpected that Primenet would assign a 100M number for P-1...

[Prime95]

Originally Posted by James Heinrich

I would find it unexpected that Primenet would assign a 100M number for P-1...

Indeed, that would be a server bug.

[Prime95]

I'm also interested in the time it takes to do the 330M-bit GCD. You're the first to attempt one with prime95.

[S00113]

Originally Posted by James Heinrich

Did PrimeNet actually assign you that number with P-1 as your preferred worktype, or did you manually add that? I would find it unexpected that Primenet would assign a 100M number for P-1...

PrimeNet assigned it to me without manual intervention. P-1 is preferred worktype for one of three workers.

I think it makes perfect sense, since this machine has more memory availiable for P-1 than most other GIMP clients. People choosing to LL test 100M numbers may not have enough memory to do P-1 themselves.

[Andi47]

Originally Posted by Andi47

I will set a worker (or maybe two) to p-1 when a flip-to-anonymous-proof version is out.

I have just downloaded 25.8 build 5 and assigned it to primenet (one thread on a machine with 2 cpus, the other cpu is running fermatsearch.exe).

With the two first-time LL-tests it assigned before I could set my preferred work-type to p-1 (and set the number of threads to 1) and the 2 LL-double-checks from 24.7 it will take until ~may until it will start the first p-1 factorization.

[petrw1]

Originally Posted by Andi47

With the two first-time LL-tests it assigned before I could set my preferred work-type to p-1 (and set the number of threads to 1) and the 2 LL-double-checks from 24.7 it will take until ~may until it will start the first p-1 factorization.

I would like to have the preferred work type set at first set-up time before it starts.

I had the same thing happen to me: I installed, answered the setup questions and it started with the default GIMPS work-types, not what I wanted to choose. So like your situation I had the choice to either unreserve these first or let them finish.

[S00113]

Originally Posted by Prime95

I'm also interested in the time it takes to do the 330M-bit GCD. You're the first to attempt one with prime95.

GCD is normally not mentioned in results.txt. Is there any way to make it, or should I run mprime with the -d option and direct the output to a file? Stage 1 was finished long ago. Stage 2 will probably finish late february/early march, or earlier if I assign something other than Fermat fatoring to the first thread, and LL-NF or D to the third to avoid other memory demanding work.