Index 1 Sequence Work for the "n^i" Aliquot Project

[EdH]

@VBCurtis: I have adjusted my script up through t40 for my smaller GPU (greater t later) and will be looking at the script further for my larger GPU, now that I finally have it working with GMP-ECM and CGBN. The larger one has 1536 cores. Thanks much!

[gd_barnes]

You'll have to let me know how long it takes you to run ECM for t40 to t45 on a ~155-160 digit number.

[EdH]

My previous post should have said t45. I haven't decided yet for t50 and am not running there.

I have a run for the c155 for 65^89 just started. It's set to do t40 and then t45. I think the GPU time would be the same for 155 and 160, but the stage 2 CPU timing would be different. Stage 2 is done on a 40-thread machine, separately.

[gd_barnes]

I wasn't clear if you were responding to my question. I didn't see anything about how much time it was for t40 to t45 for you. (Or maybe you were just beginning to test that process.)

I'm getting ~1.65-1.75 hours on my Ryzen. I got consistently ~15 mins. for t35 to t40.

Since it's a big jump, mainly I want to verify that it's worth my time to run some of them from t40 to t45 or if your process is substantially faster.

[EdH]

I don't have the best setup for real throughput, but here's a description with timings for both the c155 and the c160.

First, the overall timing is affected by the fact that if I'm running two (or more) levels, stage 2 is completed on the second machine while the next level of stage 1 is completed on the GPU. If I'm running only a single level the GPU machine has to wait for the CPU stage 2 returns.

Having described that, here are the timings for the different stages and sizes:

c155 t40 stage 1: 26:16
c155 t40 stage 2: 6:20
c155 t45 stage 1: 1:51:31
c155 t45 stage 2: 17:37

c160 t40 stage 1: 26:13
c160 t40 stage 2: 6:35
c160 t45 stage 1: 1:51:29
c160 t45 stage 2: 17:36

The timings appear the same between the two sizes. I expected this for the GPU, but am surprised it is the same for the CPU.

I've finished the c145s. I'm not sure if I'll head for the c146s just yet or do some other stuff first. I'll figure that out later.

[gd_barnes]

I'm not sure what the stages are. But even if I'm only running stage 1, I'm still a bit faster. Although your description of your processes makes me wonder if a direct comparison of that nature is reasonable.

Regardless, I feel like it's close enough that it makes sense for me to contribute for t45. I'll let you know what I do.

My jump in time was bigger than yours going from t35 to t40 vs. t40 to t45. ~15 mins vs. ~100-105 mins. About a 7x jump. Yours was ~4x. So if the jumps continued like that, it probably wouldn't make sense for me to contribute much past t45 unless I was to learn more about optimizing.

I have done nothing for optimization. Just strictly default ECM parameters. The optimization that you guys talk about goes over my head at this point.

It's interesting that your timings are about the same for different sizes but that doesn't surprise me much for such a small difference in sizes. I don't get much difference in times for those either but I do for a bigger size difference, perhaps ~10-15% for a C150 vs. a C180. You should try a bigger difference like that.

[VBCurtis]

Originally Posted by gd_barnes

I have done nothing for optimization. Just strictly default ECM parameters. The optimization that you guys talk about goes over my head at this point.

Most of my optimizations from from running ecm -v with a variety of B1 values and looking for the shortest expected time for a T-value of interest. The times hardly change within reasonable B1 values: for instance, using B1=3e6 is nearly as fast as 1e6 to complete a T35. So, I figure I may as well use B1=3e6 to spend roughly the same amount of time on T35 while also improving my chances of a 40-ish-digit factor.

On my ecm-7.01 with -v, 1e6 says 1007 curves for T35 while 3e6 says 374 curves. Both say 24.45 minutes for T35 (single-threaded).
When two times are close, like within 1%, I choose the highest B1 I can to maximize my shot at larger factors.

Ed is doing the B1 part on a GPU, and the B2 part on CPU. That's Stage 1 and Stage 2. A normal gmp-ecm invocation (or ecm.py) does both on cpu.

[EdH]

Less than 200 left!

[EdH]

I've queued these. Then I'll move back to the other thread:

33^98: 149/147
74^85: 159/147
105^95: 193/147
210^77: 180/147
552^59: 163/147

[gd_barnes]

I'll run ECM to t45 on the four C148s and the first four C149s.

[gd_barnes]

The four C148s and first four C149s have been ECM'd to t45. There was one split. These can all now be considered fully ECM'd.

Here is the list:

46^96
51^92
1155^51
12496^37
22^123
24^111
26^119*
60^87

*Split (P43-P47-P61) and can be removed.

For these "smaller" cofactors, I was getting almost exactly 1.5 hours per run. So 1 split in 12 hours. Not very good but it is a small sample size. It is still quite a bit faster than I could run NFS on any one of these.

Other previous work:

I ECM'd all of the sequences with "*"s to t45. (None are on Yoyo's reservation list.) No splits out of the 10 sequences. Overall now 1 split out of 18 for t40-t45.

I did ECM a 2nd time for t35-t40 on all sequences in this thread. It did pretty well for a 2nd pass. Overall there were 14 splits out of 215 sequences at the time: 6.5%.

At ~14-15 mins per run, it took ~50-53 hours. With 14 splits, it was one split every ~3.6-3.7 hours. Considering the average size of these monsters, it was worth it.