Links to Precompiled GMP-ECM versions

[fetofs]

Originally Posted by wblipp

Bump.

Are there 6.1 executables available for various configurations?

I've determined the 6.0.1 Windows executable I'm using on P4s and Athlon XPs fails when doing ECM with B1=260M, giving the message

GNU MP: Cannot allocate memory (size=562053120)

You can reduce memory allocation by using various -k and -treefile options (I think), and here is the binary I use for my P4 (compiled with GMP 4.2 and ECM 6.1).
I compiled it under WinXP/MinGW.

[Greenbank]

MacOS X, ECM 6.0.1 with GMP 4.2 64-bit G5 only: ecm_6.0.1_gmp_4.2_MacOSX_64bit.zip

[Mystwalker]

gmp-ecm 6.1 w/ gmp 4.2, asm-redc disabled, ntt and tuned, available here

I found this combination to be the fastest in general, but not in every situation. Let's wait for gmp-ecm 6.1.1 to do a wide-range benchmarking...

[fetofs]

Originally Posted by Mystwalker

gmp-ecm 6.1 w/ gmp 4.2, asm-redc disabled, ntt and tuned

Tuned for what?

[Mystwalker]

Originally Posted by fetofs

Tuned for what?

I tuned the ecm-params.h parameters, using the "tune" command. I found that those were some % faster on a 500 digit candidate, but (IIRC) 1% slower on a 200 digit composite.
Most other compile combinations basically yielded equal run times per curve.

[Greenbank]

Notice that every other post describes what OS and platform their binaries have been compiled for. Notice what you post lacks. Now consider what was meant by:-

Tuned for what?

[Phil MjX]

Originally Posted by Mystwalker

I tuned the ecm-params.h parameters, using the "tune" command. I found that those were some % faster on a 500 digit candidate, but (IIRC) 1% slower on a 200 digit composite.
Most other compile combinations basically yielded equal run times per curve.

Hi !

I have noticed that it is very important to tune gmp-ecm 6.1 : look for the difference for stage 2
(most of the composites are ok with both versions but in this case, and some others, it makes a great difference) :

first compilation, without tuning :

GMP-ECM 6.1 [powered by GMP 4.2.1] [ECM]
Input number is 709891330215674922888729762564179876071621230485156504316706784486089//
628388958197669654729613064737223993515869067455247541707 (126 digits)
Using B1=3000000, B2=5706890290, polynomial Dickson(6), sigma=78228681
Step 1 took 78266ms
Step 2 took 101594ms ( ! )


recompiled, adjusting ecm-params.h :

GMP-ECM 6.1 [powered by GMP 4.2.1] [ECM]
Input number is 709891330215674922888729762564179876071621230485156504316706784486089//
628388958197669654729613064737223993515869067455247541707 (126 digits)
Using B1=3000000, B2=5706890290, polynomial Dickson(6), sigma=1397049393
Step 1 took 77828ms
Step 2 took 36078ms


This may explain some disappointing results with version 6.1...

Regards.
Philippe.

PS have you noticed that running many instances of tune.exe gives differents parameters? The first trial offers me good params but I am interested in increasing the accuracy of the test : how can I change the sources to obtain a better estimation?
Thanks

[Mystwalker]

Originally Posted by Greenbank

Notice that every other post describes what OS and platform their binaries have been compiled for. Notice what you post lacks. Now consider what was meant by "Tuned for what?"

Ok, I just realized that it's not clear that my post #14 was a response to William's request in #11.
Hence, I explicitely state that it's a Windows binary. I don't know whether the optimal parameters differ across OS'es, though. If not, I've tuned it for every OS, one just needs to recompile for the resp. platform.

Originally Posted by Phil MjX

look for the difference for stage 2

What CPU do you use? Pentium M and P4 only gained a few % in my tuning tests...

PS have you noticed that running many instances of tune.exe gives differents parameters?

Yes, I also noticed this when I tested some dev-versions for David Newman. Some values deviate slightly, others from one extreme (e.g. 65535) to the other (e.g. 512).
Probably these values are not really critical or some background processes slightly modifiy the outcome of some parameters, which could result in other changed parameters.

But I don't have insight experience with this. I've only changed the values and did some benchmarking...

[wblipp]

Thanks, fetofs and Mystwalker. Those solved the problem on the P4. Anybody have a version for Athlon XP?

William

[jasong]

Since something happened that "broke" OmbooHankvald's tutorial, I'd like to request an ecm.exe file that will work well on a 2.8GHz Pentium-D with 1MB cache for each core. I don't think they're shared, but I may be wrong. I've been having some really, REALLY bad luck with my new computer(I think the d: drive is causing most, if not all, of the trouble.), so I don't have a lot of stuff at the moment, like cpu-z.

[wblipp]

Originally Posted by jasong

I'd like to request an ecm.exe file that will work well on a 2.8GHz Pentium-D with 1MB cache for each core.

Try the links in messages #12 and #14 in this thread.