New GMP!!

[VJS]

Agreed,

I did not mean to imply that increasing b2 bounds by 40% will increase your chances of finding a factor by 40% that's certain. What I did want to say is that "increasing the B2 bounds by 40% will increase the time for that curve by roughly 40%".

It just wasn't a fair comparision between the two versions with different bounds. I'm uncertain why the beta client uses larger bounds by default although I'm glad to see it doing so. I've always ran a large B2 since stage2 seems to run significantly faster than the old days of B2=100xB1, especially on dual channel boards with sufficient memory.

If both clients are running the same speeds with the same bounds I'm uncertain about the improvments. Not being harsh here or anything I just havn't followed the discussion.

[Mystwalker]

Originally Posted by Phil MjX

the --enable-asm-redc works for "smaller" numbers (<180 digits ??
the bigger ones being squarred with another method in gmp-ecm)
that's why you won't see no effect with such a "large c205 number",
try a bench with the c126 I have provided (currently ggnfs'ing)...

Actually, we don't see no effect - we see worse timings for the P4 once --enable-asm-redc is activated.

But it's true that different composites prefer different gmp-ecm variants. Yesterday, I tested the M1071.C205 with my Pentium-M.

Result:
Stage1: 6.1 --enable-asm-redc was fastest, then 6.0.1, then 6.1 - but they differ by 1.5% from first to last.
Stage2: 6.1 (both) are ~15% faster than 6.0.1


But then I put them on 55459*2^1666+1 (507 digits). Here, gmp-ecm 6.0.1 (with gmp4.2) was fastest for stage2 - by ~20%!


Seems like we either find a good way to determine which variant is the best for certain numbers, or we have to benchmark once we start ECM work on a "new" composite. I'll soon test the gwnum variant. And it really gets interesting once someone plays around with "tune".

[Mystwalker]

Originally Posted by Prime95

Have you tried the use gwnum option? GPL compatible libraries are available at ftp://mersenne.org/gimps

I just wanted to try, but sadly, I found no way doing so. gmp-ecm's help didn't give me a hint, neither did the configure's help. Did I miss something?

[Phil MjX]

Originally Posted by Mystwalker

But it's true that different composites prefer different gmp-ecm variants.

Clearly, for smaller numbers (gnfs candidates for example), v6.1.beta2 is a good choice with my pentium-m.
For larger numbers, I do agree with you that the improvement isn't so spectacular (but we have tested only two composites with a single B1=3e6 and I am sure that the developpers have reasons for their choices)!!

Let's wait for the final gmp-ecm 6.1 release .
Philippe.

[Andi47]

A few curves on M1061 (time in milliseconds) at B1=110M

			
	P95, B2=default | GMP-ECM (4.2/6.0.1), B2=100*B1	
Kurve	Step 1	Step 2 |Step 1	Step 2
1	1571005	194096 |3223343	98944
2	1570854	194047 |3248428	98974
3	1570851	193983 |3229012	99322
-----------------------------------------
average	1570903	194042 | 3233594 99080

Edit: I will do some curves with B2=4290M with GMP-ECM tomorrow.

[akruppa]

Originally Posted by Mystwalker

I just wanted to try, but sadly, I found no way doing so. gmp-ecm's help didn't give me a hint, neither did the configure's help. Did I miss something?

Say "--with-gwnum=<directory where GWNUM library is>" at configure time. Please let me know how it works.

Thanks,
Alex

[Andi47]

GMP-ECM 6.0.1 [powered by GMP 4.2] [ECM]
Input number is 2^1061-1 (320 digits)
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=4229074309
Step 1 took 3224970ms
Step 2 took 48658ms
Run 2 out of 25:
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=4059283320
Step 1 took 3233083ms
Step 2 took 48797ms
Run 3 out of 25:
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=3225355689
Step 1 took 3256811ms
Step 2 took 48855ms

I interrupted the run after 10 completed curves.

[R.D. Silverman]

Originally Posted by Andreas Schinde

GMP-ECM 6.0.1 [powered by GMP 4.2] [ECM]
Input number is 2^1061-1 (320 digits)
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=4229074309
Step 1 took 3224970ms
Step 2 took 48658ms
Run 2 out of 25:
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=4059283320
Step 1 took 3233083ms
Step 2 took 48797ms
Run 3 out of 25:
Using B1=110000000, B2=4290000000, polynomial Dickson(6), sigma=3225355689
Step 1 took 3256811ms
Step 2 took 48855ms

I interrupted the run after 10 completed curves.

I would suggest using larger step 2 limits. It is optimal to spend the
same amount of time in step 1 and step 2. This is independent of the
speed of the computer.

[PBMcL]

Originally Posted by R.D. Silverman

I would suggest using larger step 2 limits. It is optimal to spend the
same amount of time in step 1 and step 2. This is independent of the
speed of the computer.

Bob, this may be true on a one processor machine, but the disparity in memory requirements between step 1 and step 2 has led me to a different strategy for multi-processor machines in the case where step 2 is going to use most of the available RAM. Given, say, a four processor box I choose B1 and B2 so that step 2 takes just less than a quarter of the total (step 1 plus step 2) time (which may require some test curves), then I stagger start 4 copies of ECM so that only one copy at a time is in step 2. Otherwise the machine spends huge amounts of time swapping pages to/from the disk.

[akruppa]

Originally Posted by R.D. Silverman

I would suggest using larger step 2 limits. It is optimal to spend the
same amount of time in step 1 and step 2. This is independent of the
speed of the computer.

Is this correct for FFT stage 2 implementations?

Alex

[Andi47]

Originally Posted by R.D. Silverman

I would suggest using larger step 2 limits. It is optimal to spend the
same amount of time in step 1 and step 2. This is independent of the
speed of the computer.

This was just for benchmarking - to compare Prime95 with GMP 4.2/ECM6.0.1 at the same B1 and B2 bounds.