mfaktc: a CUDA program for Mersenne prefactoring

[snme2pm1]

Apparently there are new versions approaching.
In the Green camp TheJudger, in the Red (and the rest) camp BDot.
I am aware of some negative views on the Prime95 Throttle mechanism, but I do rely on having that for a notebook mounted atop a fan.
Prime95 is deployed, sometimes with throttle, when using mfaktc would be unsustainable in the current AU summer conditions.
I have to actively adjust load based on weather forecasts to prevent error exits from mfaktc.
Would a similar style of throttle facility be feasible for mfaktx?
Even better if it could track system temperature!
P.S. I am aware of discussion in another thread about faulty notebook results, yet surely better if high temperature conditions can be managed.

[James Heinrich]

Since you started factoring at 2⁶³ the CPU sieve must be used, at least for 2⁶³-2⁶⁴. I suspect that you may have "stages=0" set in mfaktc.ini which prevents the assignment from being split at the crossover point.

[Mark Rose]

Originally Posted by James Heinrich

Since you started factoring at 2⁶³ the CPU sieve must be used, at least for 2⁶³-2⁶⁴. I suspect that you may have "stages=0" set in mfaktc.ini which prevents the assignment from being split at the crossover point.

Nope. mfaktc is splitting assignments just fine above 67.

[James Heinrich]

In any case it shouldn't be an issue for most people, since the whole PrimeNet range is long since factored well beyond 2⁶⁴ -- even up to M2³² should be finished within 6 months or so. If you're redoing old TF for some reason, just split your assignments manually for now

[Mark Rose]

Originally Posted by James Heinrich

In any case it shouldn't be an issue for most people, since the whole PrimeNet range is long since factored well beyond 2⁶⁴ -- even up to M2³² should be finished within 6 months or so. If you're redoing old TF for some reason, just split your assignments manually for now

Indeed. I'm not really worried about it. It's the first time I've done work at bit levels below 68 :D

[TheJudger]

Originally Posted by Mark Rose

I found a "bug".

While double checking some trial factoring assignments in another thread, mfaktc picked 75bit_mul32 for Factor=N/A,54820379,63,66 instead of the usual barrett76_mul32_gs. It ran about 1/7th the speed that 64,66 assignments run since it was sieving on the CPU.

I think the default should be to always GPU sieve, or if the CPU sieve is to be used, to split the assignment at the cross over point.

Works as designed.
No barrett based kernel in mfaktc 0.20 is able to handle FCs below 2⁶⁴. So it isn't a matter of GPU or CPU sieving. There are only the "old" schoolbook-division kernels which can do lower FCs. In 0.20 non of these kernels is GPU sieve enabled. This will change in 0.21.
One could try to improve the automatic splitting of bitlevel, ofcourse.

Oliver

[James Heinrich]

Originally Posted by TheJudger

One could try to improve the automatic splitting of bitlevel, of course.

That's the key. I'm not sure what the current splitting logic is but if it's relatively easy it would be great to forcibly split the assignment at CPU/GPU-sieving cutoff (currently 2⁶⁴), at least if stages=1.

Certainly not a big issue considering everything (in PrimeNet at least) is long since done beyond 2⁶⁴, but I don't think it would be a bad idea to add it anyways if the code change is simple.

[lycorn]

Originally Posted by James Heinrich

everything (in PrimeNet at least) is long since done beyond 2⁶⁴

Well, not quite... There are stiil more than 70K numbers trial factored to 64 bits or less.

Granted the exponents are all < 5M, but still some people are working on them and finding many new factors . GPU sieving would be a most welcome helping hand, specially if it could be used to TF under 1M as well.

[James Heinrich]

Originally Posted by lycorn

Well, not quite... There are stiil more than 70K numbers trial factored to 64 bits or less. Granted the exponents are all < 5M...

Fair point, I stand corrected.
74759 unfactored exponents as I count them, and 319 of them are over 5M (between 5,215,801 and 5,325,181):

+----------+--------+
| count(*) | mrange |
+----------+--------+
|    16976 |      0 |
|    15886 |      1 |
|    21636 |      2 |
|    16963 |      3 |
|     2979 |      4 |
|      319 |      5 |
+----------+--------+

[Jayder]

Originally Posted by James Heinrich

Fair point, I stand corrected.
74759 unfactored exponents as I count them, and 319 of them are over 5M (between 5,215,801 and 5,325,181):

Those have been factored up to 2^65. I guess it didn't get passed to mersenne.ca.

[lycorn]

Jayder is right.
If you query Primenet directly, via Reports -> Detailed Reports -> Factoring Limits, you´ll get (as of 09:48 UTC):

61 bits - 13250
62 bits - 3257
63 bits - 53708
64 bits - 188

Total - 70403

The highest is 4699963, TFed to 63 bits.