M1061...

[clowns789]

I was doing some 1061 to 44M and left B2 blank. Is there any way to make it more efficient?

[ET_]

[Fri Oct 15 11:38:32 2004]
M1061 completed 1 ECM curves, B1=44000000, B2=4290000000
[Tue Oct 19 19:35:01 2004]
M1061 completed 24 ECM curves, B1=44000000, B2=4290000000

Luigi

[thomasn]

[Wed Oct 20 10:01:20 2004]
M1061 completed 100 ECM curves, B1=44000000, B2=4290000000

Thomas

[geoff]

Originally Posted by Xyzzy

Are we even dealing with FFTs here, like we are when we do GIMPS work?

I did a bit more testing and with mprime the graph of stage 1 time vs exponent on a 128k cache Celeron takes a big jump at about M23000, which corresponds to the crossover from a length 1024 (1K) FFT to a length 1280 FFT. On the same machine the graph for gmp-ecm is fairly smooth, at least up to M50000.

[Axel Fox]

Hmm, I have three little questions.

I think I noticed somewhere that gmp-ecm is slower in stage 1 than Prime95, but gmp-ecm is waaaaay faster in stage 2. Is this correct ?

Does the advantage of stage 2 make up for the loss in stage 1, in other words, is it better to use Prime95 or gmp-ecm.

I suspect the answer to my second question is yes and gmp-ecm, so since I want to do some ECM on a windows computer, is there a distribution of ECM for Windows, or do I have to run it in cygwin and would it still be faster than Prime95 if I do run it in cygwin ?

Ok, that last question may not be so little.

Thanks,
Axel Fox.

[geoff]

I think I noticed somewhere that gmp-ecm is slower in stage 1 than Prime95, but gmp-ecm is waaaaay faster in stage 2. Is this correct ?

Yes. Also Prime95 is limited to a stage two bound of 4.29 billion, which is OK for looking for 50 digit factors, but gmp-ecm will be needed once we start looking for 55 digit factors.

Does the advantage of stage 2 make up for the loss in stage 1, in other words, is it better to use Prime95 or gmp-ecm.

This depends on what hardware you are using. For M1061 Prime95 is faster on a P4, gmp-ecm is faster on non-sse2 machines, I am not sure about the A64.

But you can have the best of both worlds: stage one with Prime95 and then continue stage two with gmp-ecm. To do this, add the line GmpEcmHook=1 to prime.ini and change the worktodo.ini entry so that stage two is not done, then when Prime95 has finished a batch of curves, run gmp-ecm as 'ecm -resume results.txt 1 44e6-180e9' to do the stage two step on each curve.

I suspect the answer to my second question is yes and gmp-ecm, so since I want to do some ECM on a windows computer, is there a distribution of ECM for Windows, or do I have to run it in cygwin and would it still be faster than Prime95 if I do run it in cygwin ?

A Cygwin version of gmp-ecm is just as fast as any other version, provided it has been compiled correctly, but it may use a little more memory. If you have a P4 and Windows then the binary at http://www.geocities.com/greatpsycho/ is as fast as you will find anywhere. Someone else may be able to recommend binaries for other machines.

[R.D. Silverman]

Originally Posted by geoff

Yes.

<snip>

.

Allow me to add: GMP-ECM will do numbers *other* than 2^n-1 and 2^n+1.
If you want to try, e.g. 10^226+1, you must use GMP-ECM.

[Axel Fox]

Thanks for your reply geoff. After some tests it seems that my self compiled version was slightly slower on stage1 and slightly faster on stage2 than the link you gave me. (Slightly is a few seconds on a complete curve, so it didn't really make a difference).

I'm going to test the GmpEcmHook=1 thingy.

BTW : what do I do if I have completed some curves, just post it here, or mail the results to george also ? And what do my B1 and B2 have to be to do some usefull curves ?

[geoff]

Originally Posted by Axel Fox

I'm going to test the GmpEcmHook=1 thingy.

I forgot to mention a couple of common problems you might encounter if you do this. First, make sure you are using the 5.0.3 version of gmp-ecm, the so-called 5.1 version has a problem with resuming from Prime95's results.txt file. Second, some compilations (including the one I pointed you to), have a problem with DOS format text files, so you might need to convert them to plain unix format first.

BTW : what do I do if I have completed some curves, just post it here, or mail the results to george also ?

We don't want work reported to George twice, so if you post it here too then add in your post that it has already been reported to George. Perhaps Xyzzy could advise what best to do?

And what do my B1 and B2 have to be to do some usefull curves ?

Again this will depend on your hardware, since it depends on the relative efficiency of stage one vs stage two steps and also on the memory available. Keep B1=44 million fixed to suit the 50 digit level we are searching at the moment, then adjust the B2 to suit your machine. In general you should set the B2 high enough so that the stage two step takes between 50% and 100% of the stage one time.

If you use the default B2 for Prime95 (4.29 billion) or gmp-ecm (180 billion) then just report the curves as if you did them entirely on Prime95 or gmp-ecm respectively. Otherwise if you choose a non-standard B2 you will need to calculate the conversion rate yourself. There is a program in this thread http://www.mersenneforum.org/showpos...05&postcount=6 to help with this.

Of the two default B2 limits I would guess that the Prime95 one would be better suited to P4's and the gmp-ecm one to non-sse2 machines. I posted some timings earlier in this thread which suggested the best B2 was between about 10 billion and 20 billion for a P4 system. You might want to do your own timings for your setup, I would expect a larger B2 to be better on non-P4 machines. Also keep in mind that small adjustments to B2 have almost no effect on overall efficiency, so if in doubt just go with one of the defaults for simplicity.

[marc]

error404   4
thomasn  100
xyzzy    126
thomasn  100
ET        25
thomasn  100
marc     171

TOTAL    626

These are the curves which everyone has mentioned in the thread so far. So unless anyone has reported their curves to George already this is the running total as a group. I suspect xyzzy has done more than 126 so if you want to add your unmentioned curves maybe we'll be near 1000.

[thomasn]

Please add
[Fri Oct 22 00:44:29 2004]
M1061 completed 100 ECM curves, B1=44000000, B2=4290000000

to the above list !

Thomas