Simultaneous CPU and GPU Primality Testing on Windows System

[jwnutter]

Originally Posted by kriesel

Hang in there. I started climbing the gpu computing learning curve in March 2017. It takes a while.

Your RTX 2080 Super will require CUDA10-capable driver, CUDA 10 dll, and CUDA10-capable executable for mfaktc. (It varies by gpu model family; GTX10xx CUDA8)

CUDA Dlls are in the CUDA_dlls folder of https://download.mersenne.ca/

You may want to try GPU-Z as a utility on your Windows10 system to see an indication of what the computer thinks is installed for your gpu (CUDA opencl openGl etc), graphically monitor gpu parameters, maybe even log them if you want. One of many utilities listed in https://www.mersenneforum.org/showpo...74&postcount=6

Thanks!

I'll work on figuring out the how to complete these steps: CUDA10-capable driver, CUDA 10 dll, and CUDA10-capable executable for mfaktc.

But here's what GPU-Z shows for my GPU.

[jwnutter]

I think I actually got mfaktc working. I tested Factor=bla,3321932839,50,71 and mfaktc found that M3321932839 has a factor: 43654296769583327 in 2-3 seconds.

I also tested Factor=DONTpostYOURassigntmentID,110387527,73,74. It took a little under 10 minutes to complete, but I was getting a warning message (could not write checkpoint file) and nothing was saved to the results file. However, this printed in cmd when complete "no factor for M110387527 from 2^73 to 2^74 [mfaktc 0.21 barrett76_mul32_gs]"

Does this look right? Should I run other tests?

Call me crazy, but learning about all this is a ton of fun. =)

Testing Factor=DONTpostYOURassigntmentID,333999319,78,79 now. Looks like it should be done in a little under 2 hours.

[kriesel]

Originally Posted by jwnutter

Thanks!

I'll work on figuring out the how to complete these steps: CUDA10-capable driver, CUDA 10 dll, and CUDA10-capable executable for mfaktc.

But here's what GPU-Z shows for my GPU.

Attached gpu-z image looked promising. Scope out the advanced tab sometime when you have a few minutes to look in wonder that the many OpenCL parameters, etc.
Sensors tab can be fun during a short gpu run.
Get familiar with nvidia-smi command line tool at some point. That's more efficient for when you get into production mode. It's probably hiding somewhere like the following. I usually make a batch file with a short name nv.bat

:loop
"c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe"
pause
  goto loop

Your message following the one with the gpu-z image looked like you've made definite progress.
Avoid posting assignment IDs (32-characters of 0-9A-F) on the forum for any current assignment before you submit the result, or full res64 for any first primality test result before there's an independent double check reported. A moderator will probably remove that if it's too late for you to. Don't post your credit card numbers publicly either.
If the executable can't open and write to a results file or a checkpoint file, that's a permissions problem probably, that needs to be fixed.
Where did you put your working directory? These programs and C:\Program Files... don't get along well. Put the whole works in a user directory with read, write, modify, execute permissions for you.

Try mfaktc -h >>help.txt. If the program can't save to the folder, that's a problem.
Consider running at least -st to confirm it's working reliably enough to find all expected factors. (I have an igp that will find 14 out of 30 in mfakto's startup test; not good.)

Try putting the usual mfaktc invocation into a batch file to save typing on restarts. Mine are called mf.bat. I usually include redirection of all console output to a log file. Haven't hunted up a good tee program for my Win7 instances yet. Since you're on Win10 you can use its tee with the append option.

After you have the rest working well, try some ini file tuning See https://www.mersenneforum.org/showpo...postcount=3246
There's a LOT to read in that thread that grew with the development of the program.
You might find some of https://www.mersenneforum.org/showthread.php?t=23386 useful, although the Concepts post can wait; it's long and complicated.

Enjoy how solving one puzzle reveals the next. (You'll be programming in Windows batch before you're done. Resistance is futile...)

[jwnutter]

My mistake. I wasn't aware of the assignment id rule - won't happen again. What's the reasoning for this? Is it an identifier of some sort or do things get messy if another person also says they received the same assignment and it turns out the number is prime?

You hit the nail on the head. The folder was in C:\Program Files. That was the problem.

Will do on the the testing. Thanks for the advice!

Yep! I've been working on ini tuning all evening. It looks like I'm averaging about 2,850 GHz-days per day with the current settings. But I'll have to play around with things a bit more as I think one of my earlier tunes was closer to 3,000 GHz-days per day. I also need to work on a few PrintFormat revisions. The GPU is currently pulling about 220W at a load of about 88%. I'm in the process of TF'ing 150 assignments in the 333800000,74,75 range. Should be done in ~12 hours.

Absolutely, my first thought was a bat file (which I'm not great at, but can learn) to avoid some typing.

Thanks again for all your help these last few days.

I do have a 2 follow up questions for you. 1.) Is it wise to run multiple instances of mfaktc on the same GPU?, and 2.) could mfaktc be installed on a Bitmain Antminer S9 SE-16TH/s (link provided below)?



https://shop.bitmain.com/product/det...96S2W5mY1i0661

[Uncwilly]

Originally Posted by jwnutter

My mistake. I wasn't aware of the assignment id rule - won't happen again. What's the reasoning for this? Is it an identifier of some sort or do things get messy if another person also says they received the same assignment and it turns out the number is prime?

Yes. It is less important for TF. But, just don't post them to protect yourself.

[kriesel]

Originally Posted by jwnutter

1.) Is it wise to run multiple instances of mfaktc on the same GPU?, and 2.) could mfaktc be installed on a Bitmain Antminer S9 SE-16TH/s (link provided below)?

https://shop.bitmain.com/product/det...96S2W5mY1i0661

Finalize your tuning with one instance. Then evaluate how many instances is optimal for throughput and how many you want to operate routinely. (I've settled on 3 for RTX2080x and 2 for GTX1650.) Multiple instances give slightly higher sustained throughput, and keep the gpu working if one instance has a problem, runs out of work or is stopped briefly to replenish work and report results.

It appears to me the link is for an ASIC based bitcoin miner, which is unsuitable for running CUDA applications.

[jwnutter]

Ran -st and -st2 and both passed with 0 total failures.

Wow, 3 instances that seems like a lot. I agree, let me get the tuning figured out and then add instances. Just so I have something to shoot for on tuning, what kind of output are you getting on your RTX2080 when running a single instance?

Another member of my team has the same RTX2080 Super, so once I get this all figured out on my rig we'll probably do the same thing on his.

[kriesel]

Originally Posted by jwnutter

what kind of output are you getting on your RTX2080 when running a single instance?

See for full-power operation, https://www.mersenneforum.org/showpo...postcount=3246
(and the third attachment of https://www.mersenneforum.org/showpo...99&postcount=8 which graphs that)

Currently, 3 instances 2696. GhzD/day, 1 instance 2397. A complicating factor is the 3 instances are not necessarily running similar exponents and bit levels. That's while configured with half-power limit for summer:

:d0 rtx2080super 125 to 292 W, 250w nominal
"c:\Program Files\NVIDIA Corporation\NVSMI\nvidia-smi.exe" -i 0 -pl 125

GhzD/day throughput is a function of many things; gpu model, software refinements, exponent, bit level, ini file tuning, instance count, power, cooling if thermally limited, gpu and vram clock rates. Beware of overclocking that can destroy accuracy.

[jwnutter]

Wow, this is great info.

In the rtx2080super tuning document (page 1) you indicate "mfaktc v0.21 2047 enabled version". Where might I find this 2047 enabled version? The version of mfaktc v0.21 I'm currently using allows a maximum GPUSieveSize of 128.

That said, tuning is going well, however, I've not been able to push GPU load over ~89% while running a single instance.

[kriesel]

Originally Posted by jwnutter

Wow, this is great info.

In the rtx2080super tuning document (page 1) you indicate "mfaktc v0.21 2047 enabled version". Where might I find this 2047 enabled version? The version of mfaktc v0.21 I'm currently using allows a maximum GPUSieveSize of 128.

That said, tuning is going well, however, I've not been able to push GPU load over ~89% while running a single instance.

In post 10 you indicate you already downloaded the cuda 10 2047 capable version. Use that.
Ignore the ini file's comments that 128 is the max GpuSieveSize. That did not get updated. Update your own copy to state 2047 max, and start benchmarking at 256, 512, 1024, 2047. Watch your gpu load climb. But a 2048 or larger GpuSieveSIze directive in the ini file will produce an error.
(2048 x 2²⁰ =2³¹ bit position computation requires computation with unsigned 32 bit integer or larger, but the actual code is signed 32 bit integer, maximum 2³¹-1. The RTX2080x would benefit from more GpuSieveSize than the program currently supports. Unsigned 32 bit would be good at 4095 max. But there's code that uses a negative value in the same variable, that would need to be changed.)

[jwnutter]

Thanks!

Yeah, that was the problem. The comments stated max of 128 so that's as high as I had adjusted the GpuSieveSize. Problem solved when I adjusted this number to 2047. I'm currently getting about 3,300 ghz-d/day at ~98% GPU load.

That said, I'll need to start the tuning process over again, which isn't too big of a deal. At least now I can start logging several inputs/outputs in an xlsx file. I've got a few values in mind (based on your rtx2080super tuning document), but let me know if I you think of other values I should track. I see a lot of DOE in my future, when I have some free time.