GPU72 / MISFIT use for 100M digit range?

[Chuck]

Varying the GPUSievePrimes from 50000 to 100000 didn't give any improvement.

[chalsall]

Originally Posted by firejuggler

So even if we loose a bit of speed, 0.20 seem still faster (Mfaktc)

Yeah. And while we may get fewer GHzDays / Day for this work, the GHzDays Saved is far higher.

kracker has already saved some poor sole 4944.6308593750 GHzDays of wasted LLing....

[Chuck]

I've now moved up two bits and am working from 76—>77. mfaktc switched to the barrett77 kernel and the throughput has dropped another 20 to 372 Ghz-d/day.

[bcp19]

Originally Posted by Chuck

I've now moved up two bits and am working from 76—>77. mfaktc switched to the barrett77 kernel and the throughput has dropped another 20 to 372 Ghz-d/day.

Is it possible the formula to calculate work done is off?

[LaurV]

Originally Posted by Chuck

This really cuts into the normal throughput; normally I was doing 436 GHz-d/days, now only getting 394

That's normal, you can't do better by adjusting the sieve, trust me, I tried all values during my last marathon of the 332M to 71 bits. See here.
It is the 6-times-longer powering step who kicks the bucket... It was the same for v0.19 too, but as that did not report directly, but only milliseconds, it went undetected. Who did work in the range may know.

edit: sorry, didn't see all the discussion after...

[kracker]

On mfakto, something in the 60,xxx,xxx spits out 150 GHz/days, in 332,xxx,xxx it is now at 120 GHz/days.

EDIT: GPU usage at 99% both times.

[Chuck]

Originally Posted by Chuck

I've now moved up two bits and am working from 76—>77. mfaktc switched to the barrett77 kernel and the throughput has dropped another 20 to 372 Ghz-d/day.

Now working the next bit level 77—>78. Using the barrett87 kernel and throughput has fallen another 20 to 353 Ghz-d/day. I may not want to play in this part of the park after this one. Card usage has held at 99% throughout.

[Bdot]

The exponents are 29 vs. 26 bits. This means 29 vs. 26 loops of square-mod. The first 6 or 7 loops are for free (preprocessing without mod). This results in 23/20=115% or 22/19=116% of computing effort.

The length of the exponent is not part of the GHz-days calculation (so far).

This effect is in addition to the increased effort that higher-bitlevel-kernels have. Switching from one kernel to the next bigger one also causes a drop that is not part of the equation (yet).

[kracker]

I've saved 19776 GHZ/days overnight with two factors!

[Uncwilly]

Originally Posted by kracker

I've saved 19776 GHZ/days overnight with two factors!

Good stuff Maynard!

[kjaget]

Originally Posted by chalsall

Yeah... I usually get ~214, I'm now getting ~192.

Oliver... Counsel?

More data:

got assignment: exp=62989193 bit_min=71 bit_max=72 (7.59 GHz-days)
Starting trial factoring M62989193 from 2^71 to 2^72 (7.59 GHz-days)
 k_min = 18742764662160
 k_max = 37485529326194
Using GPU kernel "barrett76_mul32_gs"
[date    time]  exponent   : percent  class #, seq    |     GHz |    time |    ETA |    #FCs |      rate | SieveP. | CPU wait |
[Feb 27 12:10] M62989193   : 100.0% 4615/4620,960/960 |  214.35 |  3.188s |  0m00s |   4.06G | 1272.5M/s |   82485 |    n.a.% |
no factor for M62989193 from 2^71 to 2^72 [mfaktc 0.20 barrett76_mul32_gs]
tf(): total time spent: 51m  0.827s

got assignment: exp=332410703 bit_min=71 bit_max=72 (1.44 GHz-days)
Starting trial factoring M332410703 from 2^71 to 2^72 (1.44 GHz-days)
 k_min = 3551605314180
 k_max = 7103210637097
Using GPU kernel "barrett76_mul32_gs"
[date    time]  exponent   : percent  class #, seq    |     GHz |    time |    ETA |    #FCs |      rate | SieveP. | CPU wait |
[Feb 27 12:21] M332410703  : 100.0% 4617/4620,960/960 |  191.83 |  0.675s |  0m00s | 768.75M | 1138.9M/s |   82485 |    n.a.% |
no factor for M332410703 from 2^71 to 2^72 [mfaktc 0.20 barrett76_mul32_gs]
tf(): total time spent: 10m 48.484s

Could try rebuilding mfaktc without MORE_CLASSES defined. And maybe change to Stages=0. I'm wondering if there's some overhead in such a short time/class that's hurting performance a bit?