Ok. Will watch this thread before starting t55. I can't start on it until tomorrow night anyway, so the timing seems perfect.

Looks like I will finish t50 in a few hours. Then I'll start on 110e6.

Does anyone have a good feel for the number of curves at 110e6 for this C167?

Using heuristic of 0.31 * size, I get 51.8 digits, or a t50 plus about 5/6ths of a t50 worth of curves at 110M. gmp-ECM tells me 3583 curves at 110M is a t50, so the heuristic suggests 3000 curves.

A c167 should take about 8500 thread-hours to sieve + LA. From t50 to t51.8 is 1.8 digits of factoring at 1/50 chance per digit = 3.6% chance of a factor. Our guess of 3000 curves would take 280 hrs (at 330 sec per curve, tested on a stock i5 haswell), which is 3.3% of expected NFS time.

The chance of factor > fraction of NFS time, so 3000 curves looks about right. A bit of test-sieving on the poly you posted might refine the 8500-hr guess, but it doesn't really matter so long as we're in the neighborhood of optimal pretesting.

Is 3000 curves @11e7 the current target? Do you still want the help for that number of curves? Glad to pitch in if still needed.

[QUOTE=VBCurtis;424887]A c167 should take about 8500 thread-hours to sieve + LA..[/QUOTE]

How is the 8500 hours calculated, from prior work? I don't have the resources to complete a C167 by myself. Though I do some NFS@Home post-processing but that is pretty much fire and forget for several days.

I am passing through 250 curves @11e7. Any help would be nice since it appears this number might be a candidate for NFS@Home and the 14e queue may have a short list.

Yes; I estimated based on my experience with smaller numbers and the rough doubling of effort for every 5 digit increase in input size. If I'd thought about it, I would have used previous runs with NFS@home to get a better estimate.

A more accurate estimate would be to test-sieve a few q, find the sec/rel, and multiply by the estimated number of relations needed. This usually isn't possible for GNFS since poly select would usually happen after ECM is complete, but in this case we have your good poly so would could do so. If I were running the job myself, I'd choose among 15e/31, 15e/32, and 14e/32, but it's 14e for sure for NFS@ home.

I would guess 275M relations for 14e/32, perhaps 5% fewer for 15e due to fewer duplicate relations. Previous NFS@home runs would give you a good guess for relations needed for 14e/31, the likely parameter choice for NFS@home for C167. Previous NFS@home data should give an idea of matrix-effort required as well.

I might get to trying such a test-sieve in a day or two if nobody else does it first.

RichD - yes a GNFS 167 does seem a reasonable candidate for 14e, though there are still some proposed composites waiting for approval/queuing.

Just started ECM at 11e7. I will have three more machines become available later this week but we should hit 3000 curves by the weekend.

[QUOTE=swellman;424915]
Just started ECM at 11e7. I will have three more machines become available later this week but we should hit 3000 curves by the weekend.[/QUOTE]

I'll toss in 500 curves toward the 3000. I'll start them tonight on an old core2, will be done by Friday.

[QUOTE=RichD;424910]How is the 8500 hours calculated, from prior work? [/QUOTE]

I found logs from a C165 I ran last fall, poly score 6.67e-13. Sieve time was ~1000 hours on a 3.2 ghz core2, with 10.6M matrix taking 120 hrs to solve (on a newer 4-threaded machine). That's 4500 thread-hours on a non-HT but older machine.

Your C167 poly score is 5.6e-13, 15% lower than my C165. So, looks like my 8500hr guess was way high (at least for Core2 thread-hours); if time scales with inverse of poly score (it does for nearby scores, plus or minus 10%) we're looking at a 5000 to 5500 thread-hour project, which also means 3000 curves at 110M is a bit high. Let's end ECM at 2000 curves at 110M.

Edit- Aha! Thread-hours on a HT machine is roughly 2Ghz equivalent, while this post's data is from 3.2Ghz non-HT. That explains my estimate dropping to 5/8ths the previous one- the data is from a machine 8/5ths the speed of a fully loaded HT-haswell. Never mind, 3000 curves is fine, nothing to see here...

So when do we crack a beer? :smile:

My first machine is reporting an ETA of 164 hours for 3000 curves. Next machine is of similar performance and will become available on Thursday evening, but depending on RichD's throughput we may not need it. We can do more ECM if NFS@Home requires it, but the above calculations by VBCurtis seem solid.

The work continues...

C167
 

Passing through 510 @ 11e7.