mfaktc: a CUDA program for Mersenne prefactoring - Page 326

Mark Rose · 2022-12-30, 05:25

I've been playing with my 8 x 3070 machine tonight. It's got a lowly 2 core/2 thread G4400 CPU and single channel DDR4-2133 memory.

Compared to my older 1070s, the 3070s in this system benefit greatly from changing GPUSieveSize=128 (from 64), especially when working on 332M exponents. I saw GHz-d/d go from ~2600 to ~3050, and volatile usage in nvidia-smi go from ~75% to ~85%. CPU usage also dropped from ~60% to ~40% (system usage 40%->25%). Once running a GPUSieveSize of 128, running two instances per card resulted in equal or less overall throughput depending on CPU saturation.

Recompiling and setting GPUSieveSize=2047 gives around 3500 GHz-d/d on a 332M number (80->81 is faster than 79->80, both using barrett87_mul32_gs) and 3700 GHz-d/d for a 128M assignment (76->77). Volatile GPU usage rose to 99% and CPU usage dropped to under 10%. Guess I didn't need to buy that quad-core i5-7400 for $40.

So it turns out low GPUSieveSize is a major bottleneck on newer cards. It didn't impact my 1070s at all.

So a 40% increase in performance for a couple hours of fiddling. Not bad.

rebirther · 2022-12-30, 07:44

Quote:

Originally Posted by lalera

hi,
here is mfaktc v0.21 linux cuda12
a version for windows i can not compile
i can not test it with a rtx4090 because i have only a gtx1050ti

The app doesnt work:

Code:

./mfaktc.exe: error while loading shared libraries: libcudart.so.12: cannot open shared object file: No such file or directory

The file was smaller than the others, you need to recompile it where the lib is also included in the local folder.

kriesel · 2022-12-30, 07:58

Quote:

Originally Posted by Mark Rose

Volatile GPU usage rose to 99% and CPU usage dropped to under 10%. Guess I didn't need to buy that quad-core i5-7400 for $40.

So it turns out low GPUSieveSize is a major bottleneck on newer cards. It didn't impact my 1070s at all.

GPU-Z can show 100% usage. Try running multiple instances on a 3070. On my 2080 Super, dual instances increased throughput by 1.5%, after all conventional tuning and maxing gpusievesize to 2047 had already been done. It also looked like there would be more to be gained on a single instance if part of mfaktc was rewritten further, to unsigned int32 for gpusievesize to allow up to 4095M. Large gpusievesize also helped a 1080 and 1080Ti. https://www.mersenneforum.org/showpo...99&postcount=8
It would be interesting to see how much high gpusievesize matters on 40xx. (Anyone?)

Mark Rose · 2022-12-30, 10:02

Quote:

Originally Posted by kriesel

GPU-Z can show 100% usage. Try running multiple instances on a 3070. On my 2080 Super, dual instances increased throughput by 1.5%, after all conventional tuning and maxing gpusievesize to 2047 had already been done. It also looked like there would be more to be gained on a single instance if part of mfaktc was rewritten further, to unsigned int32 for gpusievesize to allow up to 4095M. Large gpusievesize also helped a 1080 and 1080Ti. https://www.mersenneforum.org/showpo...99&postcount=8

GPU-Z seems to be Windows only.

I did try running two instances on a single 3070 with 2047 GPUSieveSize on a 332M 80->81 assignment: throughput dropped 13% from ~3500 to ~3050.

I took a look at increasing gpu_sieve_size beyond 2047. I'm not seeing much reason why it couldn't be increased beyond that, though it probably require changing some types passed to the CUDA kernels. I haven't peeked at how it will affect CPU sieving.

Mark Rose · 2022-12-30, 10:08

Quote:

Originally Posted by kriesel

GPU-Z can show 100% usage.

And I just caught a 100%. I'm happy with 99% using a single instance.

Code:

$ nvidia-smi
Fri Dec 30 10:05:47 2022       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.108.03   Driver Version: 510.108.03   CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  Off  | 00000000:01:00.0 Off |                  N/A |
| 86%   75C    P2   235W / 270W |    424MiB /  8192MiB |    100%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  NVIDIA GeForce ...  Off  | 00000000:02:00.0 Off |                  N/A |
| 76%   67C    P2   234W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   2  NVIDIA GeForce ...  Off  | 00000000:03:00.0 Off |                  N/A |
| 83%   72C    P2   241W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   3  NVIDIA GeForce ...  Off  | 00000000:04:00.0 Off |                  N/A |
| 72%   63C    P2   240W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   4  NVIDIA GeForce ...  Off  | 00000000:08:00.0 Off |                  N/A |
| 81%   71C    P2   236W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   5  NVIDIA GeForce ...  Off  | 00000000:0B:00.0 Off |                  N/A |
| 84%   73C    P2   243W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   6  NVIDIA GeForce ...  Off  | 00000000:0C:00.0 Off |                  N/A |
| 86%   74C    P2   239W / 270W |    424MiB /  8192MiB |     99%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   7  NVIDIA GeForce ...  Off  | 00000000:0D:00.0 Off |                  N/A |
|  0%   49C    P8    26W / 270W |      3MiB /  8192MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      8757      C   ./mfaktc.exe                      421MiB |
|    1   N/A  N/A      8244      C   ./mfaktc.exe                      421MiB |
|    2   N/A  N/A      9980      C   ./mfaktc.exe                      421MiB |
|    3   N/A  N/A      8753      C   ./mfaktc.exe                      421MiB |
|    4   N/A  N/A      8801      C   ./mfaktc.exe                      421MiB |
|    5   N/A  N/A      8808      C   ./mfaktc.exe                      421MiB |
|    6   N/A  N/A      8816      C   ./mfaktc.exe                      421MiB |
+-----------------------------------------------------------------------------+

Mark Rose · 2022-12-30, 11:08

Playing around with power limits, I set them to 200 watts (down from 270, though they were using ~245) and lost only ~5% throughout. Each card behaves slightly differently.

I'll gladly save the 400 watts.

It's now 4 am. Time to sleep!

kriesel · 2022-12-30, 11:27

Quote:

Originally Posted by Mark Rose

I did try running two instances on a single 3070 with 2047 GPUSieveSize on a 332M 80->81 assignment: throughput dropped 13% from ~3500 to ~3050.

I took a look at increasing gpu_sieve_size beyond 2047. I'm not seeing much reason why it couldn't be increased beyond that, though it probably require changing some types passed to the CUDA kernels. I haven't peeked at how it will affect CPU sieving.

Performance drop is surprising.
Gpusievesize impacts an int32 variable. There is some usage in mfaktx of a negative value. So that would need to be split off to a different variable.
Or one could partially 64-bit-ize mfaktc.
If one wanted to play above exponent 2^32, a more complete 64-bit conversion would be needed, including rewriting the kernels.

lalera · 2022-12-30, 17:07

Quote:

Originally Posted by rebirther

The app doesnt work:

Code:

./mfaktc.exe: error while loading shared libraries: libcudart.so.12: cannot open shared object file: No such file or directory

The file was smaller than the others, you need to recompile it where the lib is also included in the local folder.

hi,
i have ubuntu v22.04 and cuda toolkit 12 installed
i compiled mfaktc v 0.21 and let the selftests -st and -st2 run - works
to compile it where the libs are does not work (changes nothing - i tried out with a live-cd)
i am sorry but i know not how to do this (i am not a programmer)

James Heinrich · 2022-12-30, 17:16

Quote:

Originally Posted by kriesel

It would be interesting to see how much high gpusievesize matters on 40xx. (Anyone?)

I haven't used mfaktc for years until this morning. I now have a 4090 to test on.

What would you like me to test?

Mark Rose · 2022-12-30, 19:46

Quote:

Originally Posted by James Heinrich

I haven't used mfaktc for years until this morning. I now have a 4090 to test on.

What would you like me to test?

That thing is bananas!

kriesel · 2022-12-30, 20:09

Quote:

Originally Posted by James Heinrich

I haven't used mfaktc for years until this morning. I now have a 4090 to test on.

What would you like me to test?

If you cut GpuSieveSize to 1024, how much performance do you lose?
If at GpuSieveSize=2047, you run two instances (in separate folders, very similar work; similar exponent, same bit level), do you gain combined throughput, or lose?

2022-12-30, 11:08	#3581
Mark Rose "/X\(‘-‘)/X\" Jan 2013 https://pedan.tech/ 2⁴·199 Posts	Playing around with power limits, I set them to 200 watts (down from 270, though they were using ~245) and lost only ~5% throughout. Each card behaves slightly differently. I'll gladly save the 400 watts. It's now 4 am. Time to sleep! Last fiddled with by Mark Rose on 2022-12-30 at 11:08

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1724	2023-06-04 23:31
gr-mfaktc: a CUDA program for generalized repunits prefactoring	MrRepunit	GPU Computing	42	2022-12-18 05:59
The P-1 factoring CUDA program	firejuggler	GPU Computing	753	2020-12-12 18:07
mfaktc 0.21 - CUDA runtime wrong	keisentraut	Software	2	2020-08-18 07:03
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51

2022-12-30, 05:25	#3576
Mark Rose "/X\(‘-‘)/X\" Jan 2013 https://pedan.tech/ 2⁴×199 Posts	I've been playing with my 8 x 3070 machine tonight. It's got a lowly 2 core/2 thread G4400 CPU and single channel DDR4-2133 memory. Compared to my older 1070s, the 3070s in this system benefit greatly from changing GPUSieveSize=128 (from 64), especially when working on 332M exponents. I saw GHz-d/d go from ~2600 to ~3050, and volatile usage in nvidia-smi go from ~75% to ~85%. CPU usage also dropped from ~60% to ~40% (system usage 40%->25%). Once running a GPUSieveSize of 128, running two instances per card resulted in equal or less overall throughput depending on CPU saturation. Recompiling and setting GPUSieveSize=2047 gives around 3500 GHz-d/d on a 332M number (80->81 is faster than 79->80, both using barrett87_mul32_gs) and 3700 GHz-d/d for a 128M assignment (76->77). Volatile GPU usage rose to 99% and CPU usage dropped to under 10%. Guess I didn't need to buy that quad-core i5-7400 for $40. So it turns out low GPUSieveSize is a major bottleneck on newer cards. It didn't impact my 1070s at all. So a 40% increase in performance for a couple hours of fiddling. Not bad.