P-1 factoring anyone? - Page 43

petrw1 · 2011-03-25, 16:42

Quote:

Originally Posted by S34960zz

The TF limits may also need re-weighting to allow for the huge change in TF productivity due to GPU computing contributions.

Agreed....

OR if the server can tell if a PC requesting TF is using the GPU the server could assign it an extra bit of TF just ahead of the LL wavefront.

Rodrigo · 2011-03-25, 17:07

Quote:

Originally Posted by petrw1

If your worktype is TF you will get assignments in the 80M range.
If TF-LMH; then in the 180M range.

petrw1,

Hmm... the "Worker Windows" setting for that P233 that's doing TF in the (now) 180M range is "First-time tests." I assure you, I haven't requested (and didn't even know that I was doing) any work in the LMH area!

I imagine that the automatic assignment to that range has to do with the limited capabilities of that processor. So could I say that my puny P233 is doing work "beyond the wavefront"?

Rodrigo

James Heinrich · 2011-03-25, 18:13

Quote:

Originally Posted by Rodrigo

setting for that P233 that's doing TF in the (now) 180M range is "First-time tests."

You asked the server for first-time LL work, but the server looked at what kind of system it is and said "no, you're too slow for that" and gave you what it thought was the most suitable kind of work for your CPU.

There's two sets of automatically-assigned TF work: TF-LMH will do trial factoring up to low limits only (2^65 or so), which (due to limited supply of unfactored exponents) is now out around 180M. Regular TF works a little bit ahead of the first-time LL tests and bring the exponents up to a higher bit depth (2^70 or so). After that the P-1 work is done, then LL, then finally double-check LL. TF-LMH is way ahead of the first-time LL (+130M); TF is well ahead of LL (+30M), and P-1 is struggling to keep in front of LL (in fact it's not, and many P-1 are of necessity done by LL workers, many times not as well as they should be).

Rodrigo · 2011-03-25, 23:06

Thanks, James.

I'd thought that in order to do LMH work one needed to volunteer specifically for it, but I guess this is one circumstance where things work differently.

My understanding of GIMPS (and PrimeNet) gets better every week, thanks to folks like you and petrw1 who patiently explain things. Much appreciated!

Rodrigo

petrw1 · 2011-03-26, 15:08

Quote:

Originally Posted by ckdo

Perhaps the "do last bit of TF after P-1" change should be undone in order to effectively halve that gap within a foreseeable timeframe.

It appears they may have heard you.
There are over 19,000 new TF assignments available in the 54M range now.

And I have 3 of them - all to 2^69; the last bit.

And if this continues you will have that much more time to finish the 90M range

petrw1 · 2011-03-30, 22:18

By moving P-1 up to the 53M range there are now over 15,000 LL tests ready in the 53M range.

I'm curious to see if that is enough of a head start to keep P-1 from holding up LL for a while.

LL is getting close to the 53M range.

nucleon · 2011-04-05, 07:54

Any GPU P-1 code out there?

-- Craig

lorgix · 2011-04-05, 08:09

If there is, I am so getting a new GPU.

ET_ · 2011-04-05, 09:58

Quote:

Originally Posted by nucleon

Any GPU P-1 code out there?

-- Craig

Give me a pointer to a (somewhat easy) P-1 implementation, and I'll try to recode it for a GPU...

(as long as George is not working on it already)

Luigi

petrw1 · 2011-04-05, 16:25

I just did a little analysis...I did the LL Results report for the last 13 days.

I only counted first time LL tests in the 37-53M ranges and excluded double checks. 42% were in the 5xM range.
Over those 13 days the average was just over 180 LL completions per day.

I know of no way to similarly count P-1 completions in the required range; nor to determine which are P-1 only and which are P-1 required before the LL tests for the same PC on the same exponent.

In the last Recent Results (46 minutes worth) there were 8 P-1 completions in the 4x-5xM ranges but that is a pretty small sample.

With a current LL "capacity" of 180 per day the project needs at least as many P-1 completions in the current LL range to keep up. Actually about 6% more than that to allow for those that find a factor.

James Heinrich · 2011-04-05, 16:39

So, at roughly 4.15GHz-days per P-1, we need (4.15 * 180 * 1.065) = 796GHz-days/day of P-1 power to keep up with LL.

2011-03-30, 22:18	#468
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 3·5²·71 Posts	A milestone of sorts.... By moving P-1 up to the 53M range there are now over 15,000 LL tests ready in the 53M range. I'm curious to see if that is enough of a head start to keep P-1 from holding up LL for a while. LL is getting close to the 53M range.

2011-03-25, 23:06	#466
Rodrigo Jun 2010 Pennsylvania 947 Posts	Thanks, James. I'd thought that in order to do LMH work one needed to volunteer specifically for it, but I guess this is one circumstance where things work differently. My understanding of GIMPS (and PrimeNet) gets better every week, thanks to folks like you and petrw1 who patiently explain things. Much appreciated! Rodrigo

2011-04-05, 07:54	#469
nucleon Mar 2003 Melbourne 5·103 Posts	Any GPU P-1 code out there? -- Craig

2011-04-05, 08:09	#470
lorgix Sep 2010 Scandinavia 3×5×41 Posts	If there is, I am so getting a new GPU.

2011-04-05, 16:25	#472
petrw1 1976 Toyota Corona years forever! "Wayne" Nov 2006 Saskatchewan, Canada 3·5²·71 Posts	I just did a little analysis...I did the LL Results report for the last 13 days. I only counted first time LL tests in the 37-53M ranges and excluded double checks. 42% were in the 5xM range. Over those 13 days the average was just over 180 LL completions per day. I know of no way to similarly count P-1 completions in the required range; nor to determine which are P-1 only and which are P-1 required before the LL tests for the same PC on the same exponent. In the last Recent Results (46 minutes worth) there were 8 P-1 completions in the 4x-5xM ranges but that is a pretty small sample. With a current LL "capacity" of 180 per day the project needs at least as many P-1 completions in the current LL range to keep up. Actually about 6% more than that to allow for those that find a factor.

2011-04-05, 16:39	#473
James Heinrich "James Heinrich" May 2004 ex-Northern Ontario 1000010110101₂ Posts	So, at roughly 4.15GHz-days per P-1, we need (4.15 * 180 * 1.065) = 796GHz-days/day of P-1 power to keep up with LL.