Let's finish PRP-CF DC's.

[kruoli]

You are sincerly welcome!

The machine is an AMD Threadripper 1950X. In the moment, I am running two threads per worker.

Previously, with all memory slots populated, I had this:

Timings for 480K FFT length (16 cores, 1 worker):  0.60 ms.  Throughput: 1661.77 iter/sec.
Timings for 480K FFT length (16 cores, 2 workers):  0.63,  0.56 ms.  Throughput: 3369.52 iter/sec.
Timings for 480K FFT length (16 cores, 4 workers):  0.96,  0.88,  0.94,  1.01 ms.  Throughput: 4239.61 iter/sec.
Timings for 480K FFT length (16 cores, 8 workers):  1.70,  1.71,  1.89,  1.89,  1.90,  1.90,  1.77,  1.70 ms.  Throughput: 4438.59 iter/sec.
Timings for 480K FFT length (16 cores, 16 workers):  3.40,  3.42,  3.44,  3.45,  3.36,  3.40,  3.42,  3.43,  3.46,  3.52,  3.43,  3.47,  3.43,  3.51,  3.48,  3.43 ms.  Throughput: 4651.21 iter/sec.

So I ran only one thread per worker. I guess that the combination of small FFT with the big CPU caches made these results possible.

Now, with only two memory modules, I have (attention, slightly bigger FFT size!):

Timings for 512K FFT length (16 cores, 1 worker):  0.65 ms.  Throughput: 1528.24 iter/sec.
Timings for 512K FFT length (16 cores, 2 workers):  0.52,  0.55 ms.  Throughput: 3724.46 iter/sec.
Timings for 512K FFT length (16 cores, 4 workers):  0.88,  0.90,  0.91,  0.93 ms.  Throughput: 4419.46 iter/sec.
Timings for 512K FFT length (16 cores, 8 workers):  1.76,  1.77,  1.76,  1.82,  1.75,  1.77,  1.81,  1.85 ms.  Throughput: 4480.33 iter/sec.
Timings for 512K FFT length (16 cores, 16 workers):  6.58,  6.58,  6.65,  6.74,  6.57,  6.56,  6.56,  6.86,  6.72,  6.70,  6.65,  6.84,  6.98,  6.58,  6.56,  6.70 ms.  Throughput: 2397.05 iter/sec.

In the moment, I don't want to tinker too much with that machine. Since last week, I never got it up running with four memory channels. I already cleaned the machine and reseated the CPU two times. I used a totally different set of DIMMs and also permutated the order in which I installed them.

In contrast, on an AMD Ryzen 3800X, I got this (with larger FFTs, again):

Timings for 560K FFT length (8 cores, 1 worker):  0.35 ms.  Throughput: 2826.81 iter/sec.
Timings for 560K FFT length (8 cores, 2 workers):  0.51,  0.50 ms.  Throughput: 3991.92 iter/sec.
Timings for 560K FFT length (8 cores, 4 workers):  0.92,  0.94,  0.94,  0.93 ms.  Throughput: 4279.59 iter/sec.
Timings for 560K FFT length (8 cores, 8 workers):  2.09,  2.08,  2.07,  2.04,  2.07,  2.17,  2.06,  2.06 ms.  Throughput: 3847.43 iter/sec.
Timings for 560K FFT length (8 cores hyperthreaded, 1 worker):  0.36 ms.  Throughput: 2758.59 iter/sec.
Timings for 560K FFT length (8 cores hyperthreaded, 2 workers):  0.48,  0.48 ms.  Throughput: 4141.37 iter/sec.
Timings for 560K FFT length (8 cores hyperthreaded, 4 workers):  0.93,  0.88,  0.90,  0.91 ms.  Throughput: 4430.35 iter/sec.
Timings for 560K FFT length (8 cores hyperthreaded, 8 workers):  2.18,  2.28,  2.13,  2.12,  2.09,  2.21,  2.29,  2.26 ms.  Throughput: 3647.35 iter/sec.

It runs all four DIMM slots populated, running at 3,600 MHz, 16-16-16-36. Interestingly, 8 cores hyperthreaded with 4 workers is the fastest setting. It is around the same speed as the 1950X with my current settings (with somewhat smaller FFTs), which has double the cores. I did not find my old benchmarks of hyperthreading on the 1950X, but IIRC, it was not beneficial for Prime95.

[Uncwilly]

Fancy graphic to gawk at.

[LaurV]

They'll finish faster (when red reached the point where green was when we began, and not where red will intersect green, that's because new prps don't need dc anymore - maybe sometime towards the end of February?)

[Uncwilly]

There are/were still (I think) some of the first time CF-PRP's being turned in that are not being done by v29. The rate of change will accelerate quickly once we do get close to those that were turned in shortly after v30 came out.

[kruoli]

Currently, in the range from 9M to 11M, there are 2,049 PRP-CF-DCs assigned. In total, there are 26,114 exponents to go.

That means, we have a progress of around 485 exponents per day, so ETA with our current speed is around 54 days.

PS: What will happen, when we run out of PRP-CF-DC work? Will Prime95 fetch PRP-CF instead?

[Uncwilly]

I believe that it would roll over to FT PRP-CF.

[Uncwilly]

Originally Posted by kruoli

Yesterday, a machine with 16 workers of mine all doing PRP-CF-DC broke down and spontaneously decided to ignore two of its memory channels.

I have a 6 core machine that I am using to do LL-DC and PRP-CF-DC (3 cores to each task.) I had the 3 cores for PRP-CF-DC as 1 worker. After seeing your post I shifted to 3 workers 1 core each for a while. It turns out I get better throughput by having a single worker. Once we get caught up to the PRP-CF-FTC the worker will be reverted to LL-DC work.

[kruoli]

The wavefront will likely hit 10M today!

Currently, in the range from 9M to 11M, there are 1,544 PRP-CF-DCs assigned. In total, there are 22,497 exponents to go.

That means, we have a progress of around 517 exponents per day, so ETA with our current speed is around 44 days.

[Uncwilly]

Originally Posted by kruoli

The wavefront will likely hit 10M today!

Not quite yet.
9997987 is the highest DC
10005887 is a FTC as are: 10144691 and 10005887

[kruoli]

Sorry, bad phrasing:
I wanted to say that we will likely hit 10M in the assigment wavefront today (when using the time zone of this forum).

[LaurV]

Originally Posted by kruoli

so ETA with our current speed is around 44 days.

Which matches well my initial approximation above, for end of February. Looking to my status page (which shows the number of people doing each kind of work, each one with his own, but the number of participants is the same if you access yours) I see new people "jumping in" daily, so the given time is quite realistic.