P-1 factoring anyone?

[Mr. P-1]

Originally Posted by petrw1

There seems to be a general consensus that Stage 2 takes about twice as long as Stage 1 for P-1.

I haven't actually measured it, but I would guess it takes about 30-50% longer for me. It's certainly nowhere near twice as long.

Obviously the amount and speed of memory will account for a great deal of the difference between different machines.

[Mr. P-1]

Originally Posted by Kevin

No, they're related.

Ultimately, throughput is maximized if every system does the task it's (relatively) the best at compared to other machines. If a modern computer has a ratio of 1:1 between time for a P-1 test and time for a TF test, and my computer has a ratio of 10:1, I'd be greatly under-utilizing my system by doing P-1 testing. However, since there is a shortage of P-1 testing, I'm willing to "under-utilize" my system to a reasonable degree if it puts CPU time where it's needed more. So the lack of P-1 testing is just saying that I have a looser standard for what is "considerably less" efficient.

Bear in mind that what your system does influences what other systems do. If you do P-1, you take that task from another machine, ultimately the one that gets the LL assignment. Even a cursory examination of the factoring limits behind the leading edge reveals that a substantial proportion of them are not getting any stage 2. If you take on this type if work, then you have a chance of finding factors that the project would otherwise miss.

On the other hand TF is well advanced, so if you do this kind of work, you'll just advance it further without relieving other systems of work they're not well suited to, or finding factors they would miss.

[James Heinrich]

Originally Posted by Kevin

Is a PIII@1.1 GHZ with 400MB assigned (out of 512MB) powerful enough to be worth putting on P-1? It's a server box that's more or less dedicated to crunching until I have the time to make Samba work, so not very much system overhead or concern for system responsiveness.

Backing up this debate with some numbers, this is what my credit-efficiency calculator has to say about your PentiumIII:
http://mersenne-aries.sili.net/throu...essor&mhz=1100

TF (up to 2^63) on the PentiumIII is vastly superior to FFT performance (by 4x-10x). TF >= 2^64 isn't particularly efficient either, so at that point P-1 would probably be (currently) more GIMPS-beneficial.

Note: I'm not sure if my calculator is flawed, or if PrimeNet is really giving an abnormally-large amount of credit for TF compared to FFT... (updates in my calculator thread, when I have any).

[cheesehead]

Originally Posted by Mr. P-1

Bear in mind that what your system does influences what other systems do. If you do P-1, you take that task from another machine, ultimately the one that gets the LL assignment. Even a cursory examination of the factoring limits behind the leading edge reveals that a substantial proportion of them are not getting any stage 2. If you take on this type if work, then you have a chance of finding factors that the project would otherwise miss.

On the other hand TF is well advanced, so if you do this kind of work, you'll just advance it further without relieving other systems of work they're not well suited to, or finding factors they would miss.

Thanks for a clear explanation of something I struggled to convey earlier. :-)

[cheesehead]

Originally Posted by James Heinrich

Backing up this debate with some numbers, this is what my credit-efficiency calculator has to say about your PentiumIII:
http://mersenne-aries.sili.net/throughput.php?cpu=Intel(R)+Pentium(R)+III+processor&mhz=1100

It said:

ERROR: Could not find benchmark data for "Intel(R) Pentium(R) III processor"

when it spoke to me just now after I clicked on your link.

When I manually selected "Intel Pentium III" or "Intel(R) Pentium(R) III processor" (what's the difference?), it worked fine, though.

Wait! I see the difference: the single benchmark available for "Intel Pentium III" results in FFT throughput % Eff. of 94.4 to 100 (but no info on TF).

whereas the 4 benchmarks for "Intel(R) Pentium(R) III processor" result in miserably-red-backed FFT % Eff. of only 8.8 to 14.4 but cheerily-green (I think) % Eff. of 99.7 to 100 for TF at 2^61 or below.

See? That's the Genuine "Intel(R)" difference. Wow!

[mdettweiler]

Originally Posted by cheesehead

It says:

ERROR: Could not find benchmark data for "Intel(R) Pentium(R) III processor"

when it speaks to me, anyway.

Same here.

[James Heinrich]

Sorry, I changed the CPU links to include L2/L3 cache data (to differentiate same-named CPUs with different amounts of cache). And I can't edit my posts anymore to change the links to point to the correct CPU This is the updated link:
http://mersenne-aries.sili.net/throu...|256|0&mhz=800

And yes, some of my (especially older) data has FFT benchmarks only, no TF, so having nothing to compare it to it shows the relative performance of the various FFT sizes (which are all reasonably close to each other), but once you throw TF in the mix the TF numbers are very much higher (still trying to figure if this is my bug or a PrimeNet weirdness) so the FFT-based work looks like a generally bad idea.

[garo]

I don't think your benchmarks are wrong. PIII really is that much better at factoring under 63 bits. I get similar results with my Athlon.

[Kevin]

Originally Posted by Mr. P-1

Bear in mind that what your system does influences what other systems do. If you do P-1, you take that task from another machine, ultimately the one that gets the LL assignment. Even a cursory examination of the factoring limits behind the leading edge reveals that a substantial proportion of them are not getting any stage 2. If you take on this type if work, then you have a chance of finding factors that the project would otherwise miss.

On the other hand TF is well advanced, so if you do this kind of work, you'll just advance it further without relieving other systems of work they're not well suited to, or finding factors they would miss.

Yes, I'm aware of all of that. It doesn't change the fact I've been repeating from the start that I don't want to do P-1 testing if it will effectively turn my 1100MHZ processor into a 100 MHZ processor.

[James Heinrich]

Originally Posted by garo

I don't think your benchmarks are wrong. PIII really is that much better at factoring under 63 bits. I get similar results with my Athlon.

I'm still fine-tuning the TF part of my throughput calculator, but the numbers look a little better now -- for a PIII both FFT-based work and TF>=2^64 drop to around 1/3 the speed of TF<=2^62.

Originally Posted by Kevin

I don't want to do P-1 testing if it will effectively turn my 1100MHZ processor into a 100 MHZ processor.

It's not quite that bad, but your PIII @ 1.1GHz performs roughly equivalent to:
* TF <= 2^62: Core2Duo @ 600MHz
* TF >= 2^64: Core2Duo @ 225MHz
* FFT (any size): Core2Duo @ 200MHz

So, since it's a matter of triple the throughput TF to 2^63 (or less) is most desirable. If you're doing TF on 2^64 or higher, you may as well be doing P-1, the both (approximately) equally (in)efficient in terms of relative performance.

[Kevin]

Thanks, those are the kind of numbers I was looking for. Even in LMH ranges, everything has been done to 2^63 (except for 90-91M, which is in progress), and I wasn't going to bother with that anyways. I'll keep it on P-1 unless the memory usage impedes system performance, or it turns out the tests take a lot longer than they should.