P-1 factoring anyone?

[lfm]

OK, I'll just let it grind on a few 50M P-1 units for now then. I guess what misled me most was the primenet report indicating LOTS of "available" P-1 work units in the 45M-49M range. It just refused to let me work on them. I think the first time LL units we are doing get to do those P-1 steps anyway and that doesn't seem to be too big a problem in the end.

As for being an old timer, I hesitate to mention I have actually been around quite a while, I did a few LL tests less than 1 million! I just haven't learned that much I guess! grin.

[cheesehead]

Originally Posted by lfm

As for being an old timer, I hesitate to mention I have actually been around quite a while, I did a few LL tests less than 1 million! I just haven't learned that much I guess! grin.

We're each following our own individual paths through space-time, and it's all relative.

- - -

If we had PrimeNet assigning P-1 (and last-bit TF) in the 45M-49M range, then there'd be the turnaround time for the P-1 assignee to complete and report the result, then more turnaround time for the last bit(s) TF assignee to complete and report that result, before the exponent again became available for L-L assignment. This could easily result in large numbers of 45M-49M exponents being unavailable for L-L assignment for a while.

- - -

Meanwhile, my guess is that most L-L assignees do not skip the P-1 and last-bit TF that come packaged with their assignment; the skippers are only a minority. So the P-1/last-bit TF/LL separation is only a minor optimization.

In essence, we're catering to the LL-only bunch in order to keep them happily contributing to GIMPS without having to do (or skip) the part they don't like. By having willing volunteers do the preliminary factoring in an optimal way, we avoid having GIMPS progress slowed down by those who don't want to do that part.

[Primeinator]

Once my main computer is no longer available to test 24/7 on LLs, I may consider doing P-1 if it is this badly needed. What is the average time for a P-1 (both stages) on a 2.0Ghz Dual core in the 50M range (or wherever the P-1 Stage 1/Stage 2 is needed most)? However, I am also considering assisting the 100M Digit Prefactor Project.

[Mr. P-1]

Originally Posted by Primeinator

Once my main computer is no longer available to test 24/7 on LLs, I may consider doing P-1 if it is this badly needed. What is the average time for a P-1 (both stages) on a 2.0Ghz Dual core in the 50M range (or wherever the P-1 Stage 1/Stage 2 is needed most)? However, I am also considering assisting the 100M Digit Prefactor Project.

Probably around a day or so per core.

[petrw1]

Originally Posted by Mr. P-1

Probably around a day or so per core.

I have a Dual COre 2.33 and with 800MB allocated it takes about 50 hours for 1 core for P-1

[S00113]

Originally Posted by Kevin

What difference does it make if that trial factoring is done now or in 5 years?

It is a waste of CPU cycles.

64bit CPUs (with 64bit OS) can do low level factoring four times faster than 32bit CPUs. There is almost no difference in LL testing speed. This means that 3/4 of the time spent factoring to low limits using a 32bit CPU is wasted, because the project would do four times more useful work if those 32bit CPUs was LL testing and a few 64-bit CPUs did all the trial factoring to low limits.

In about ten years we will probably see the first 128bit CPUs, wich will be four times more efficient at factoring to a few bits higher limits. If we keep the leading edge of factoring just ahead of LL, we can do those ranges in 1/4 of the CPU cycles then.

[garo]

Originally Posted by S00113

It is a waste of CPU cycles.
<snip>

Good point on 32 vs. 64 bits. I admit to doing some lower level factoring on 60M, 80M, 94M and 192M ranges but my excuse is that I have a 64-bit Opteron that does this work really fast and is relatively crap at everything else. I do not do any LMH type activity on 32-bit machines except real snails like a recently retired PIII 600MHz belonging to my parents that worked well as a space heater.

[petrw1]

Originally Posted by S00113

64bit CPUs (with 64bit OS) can do low level factoring four times faster than 32bit CPUs. There is almost no difference in LL testing speed.

Unless I misunderstand you this does not match my observations:

My PIII 866 can do <64 Bit LMH at the rate of 0.6 Ghz Days per day.
Over 64 Bits it drops to 0.2 per day.
Similarly a 22M DC (18 Ghz-Days) is "Time'd" at 90 days or again 0.2 per day.

My Duron (1.3 Ghz) gets almost 1 per day at LMH and 0.4 at TF above 64 bits.

My PIV 2.8 can only do 1.2 per day at LL.

This tells me my PIII and Duron are really only productive (and I use the term loosely) doing LMH.

[James Heinrich]

Originally Posted by petrw1

This tells me my PIII and Duron are really only productive (and I use the term loosely) doing LMH.

I concur. I have a P3-class Celeron that I'm using only for TF to 2^64 because it's pathetically inefficient at all other types of work: 3x lower relative throughput at TF>2^64; 4x lower relative throughput at current FFT sizes:
http://mersenne-aries.sili.net/throu...256|0&mhz=1000

An 800MHz Duron shows nearly-identical relative performance:
http://mersenne-aries.sili.net/throu...r|64|0&mhz=800

[Primeinator]

Originally Posted by S00113

It is a waste of CPU cycles.

64bit CPUs (with 64bit OS) can do low level factoring four times faster than 32bit CPUs. There is almost no difference in LL testing speed. This means that 3/4 of the time spent factoring to low limits using a 32bit CPU is wasted, because the project would do four times more useful work if those 32bit CPUs was LL testing and a few 64-bit CPUs did all the trial factoring to low limits.

In about ten years we will probably see the first 128bit CPUs, wich will be four times more efficient at factoring to a few bits higher limits. If we keep the leading edge of factoring just ahead of LL, we can do those ranges in 1/4 of the CPU cycles then.

How is it four times faster and what constitutes "low" bit levels? Is it not four times faster at higher bit levels? And 128-bit CPUs already exist... if you want to fork out the money to buy a server CPU.

[petrw1]

Originally Posted by Primeinator

How is it four times faster and what constitutes "low" bit levels?

It has something to do with 32 bit registers. Above 64 bits the software has to use double the registers for each operation.