GPU LL Testing FAQ

[Luis]

Originally Posted by kriesel

https://www.primegrid.com/forum_thre...ap=true#123421 seems to say it is an application developer's issue. And relates to OpenCL performance.
But, have you considered going dual-boot as a workaround?

Yeah, I know it concerns OpenCL, but I can't upgrade nvidia drivers if I want AP project to work at its best. I believe BOINC projects use their own CUDA libraries. For example GPUGRID project directory contains libcudart.so.8.0 and libcufft.so.8.0.


Ubuntu 18.04 is already installed. Windows 10 too. It's just that I have a lot of working stuff on the current OS. BOINC, torrents, configurations, etc...

[Flammo]

Hi forum.

I'm new to CudaLucas, and last night I managed to LL an 89 mil exponent in 4 hours on a GTX 750 Ti. I uploaded the result to primenet and it says it could be prime. Then this morning I switched from CUDA4.2 to 8.0 after I discovered where to download the libraries from, and now an exponent will take a month to compute, not 4 hours.

It's clear to me that what I uploaded using CUDA4.2 was skipping over everything way too quickly, no idea what it's doing, but it's not a useful upload.

What went wrong, and what should I do about the bad result I uploaded?

Thanks in advance.

[kriesel]

Originally Posted by Flammo

Hi forum.

I'm new to CudaLucas, and last night I managed to LL an 89 mil exponent in 4 hours on a GTX 750 Ti. I uploaded the result to primenet and it says it could be prime. Then this morning I switched from CUDA4.2 to 8.0 after I discovered where to download the libraries from, and now an exponent will take a month to compute, not 4 hours.

It's clear to me that what I uploaded using CUDA4.2 was skipping over everything way too quickly, no idea what it's doing, but it's not a useful upload.

What went wrong, and what should I do about the bad result I uploaded?

Thanks in advance.

Actually you didn't LL test an 89M exponent in 4 hours. That's not possible with even the fastest existing gpu model. What happened is you stumbled into a case like one of the first several known bugs for CUDALucas, which you can read about in the attachment on the bug and wish list post in the CUDALucas-specific reference thread at https://www.mersenneforum.org/forumdisplay.php?f=154 There's also an updated readme available in that CUDALucas-specific reference thread.

If you want, you could PM Madpoo the exponent and the request to remove the entry. He may have already spotted the false prime report. Generally, CUDALucas v2.05.1 or earlier version prime reports are not taken very seriously because it's known to fail this way; the same thing that makes it generate a continuous repeating stream of zero residues makes it run implausibly fast. (I speculate that certain subroutine calls fail, so they take no time, or that data transfers fail, so iterations finish far too quickly.) Early on, I had a couple of occurrences like that.

I suggest that if this was an exponent assigned to you, you rerun it after getting your CUDALucas version, ini file settings, and fft benchmarking and thread benchmarking and memory test done, and successfully run the selftest using a variety of fft lengths up to 8M, and maybe a small known prime mersenne repeat too. (Don't bother to submit that, it won't be accepted.)

Running gpu GIMPS applications is not like running prime95. It requires more care and attention, especially while learning it and setting it up with a correctly working configuration. I redirect all console output to log files, so I can review everything later.

It's possible you've stumbled on a case I've not seen. Please assemble the details and send them to me if it is a different case than the ones already listed in the CUDALucas bug and wish list table. Exact CUDALucas version, OS, exponent, sample console output, gpu, etc. If you don't have console output but do have an interim save file, you may be able to recreate some by rerunning from the last save. I'll add a confirmed new case to the list.

[Flammo]

Thanks Kriesel for the detailed reply.

Yes there was an element of sarcasm to the 4 hour LL test :)

I just checked up on my results and see that it's already disappeared so it seems to have been dealt with. I've now got CudaLucas setup and it seems to be working properly now.

I somehow didn't notice the reference thread sticky before, but I'll make sure to spend the time to read through it all and see what else I can learn and do, to check it's all running correctly.

Good idea to redirect console output into log files, I'll be sure to check out how to do that too.

Thanks

[kriesel]

Originally Posted by Flammo

Good idea to redirect console output into log files, I'll be sure to check out how to do that too.

Thanks

I use a batch file and shortcuts. It can be pretty simple. Something like this, called cl.bat

echo batch wrapper reports (re)launch at %date% %time% >>cl.txt
CUDALucas2.06beta-CUDA5.0-Windows-x64.exe -d 1 >>cl.txt
echo batch wrapper reports exit at %date% %time% >>cl.txt

[Luis]

It worked. Finally I have CUDALucas 2.06Beta.

luis18@luis18:~/cudalucas-code-r102-trunk$ lsb_release -a
No LSB modules are available.
Distributor ID:    Ubuntu
Description:    Ubuntu 18.04.2 LTS
Release:    18.04
Codename:    bionic
luis18@luis18:~/cudalucas-code-r102-trunk$ uname -r
4.15.0-47-generic
luis18@luis18:~/cudalucas-code-r102-trunk$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85
luis18@luis18:~/cudalucas-code-r102-trunk$ ./CUDALucas -v

CUDALucas v2.06beta

luis18@luis18:~/cudalucas-code-r102-trunk$ ldd ./CUDALucas
    linux-vdso.so.1 (0x00007ffd11dfc000)
    libcufft.so.9.1 => /usr/lib/x86_64-linux-gnu/libcufft.so.9.1 (0x00007f8879f79000)
    libcudart.so.9.1 => /usr/lib/x86_64-linux-gnu/libcudart.so.9.1 (0x00007f8879d0b000)
    libm.so.6 => /lib/x86_64-linux-gnu/libm.so.6 (0x00007f887996d000)
    libnvidia-ml.so.1 => /usr/lib/x86_64-linux-gnu/libnvidia-ml.so.1 (0x00007f8879344000)
    libc.so.6 => /lib/x86_64-linux-gnu/libc.so.6 (0x00007f8878f53000)
    libdl.so.2 => /lib/x86_64-linux-gnu/libdl.so.2 (0x00007f8878d4f000)
    libpthread.so.0 => /lib/x86_64-linux-gnu/libpthread.so.0 (0x00007f8878b30000)
    librt.so.1 => /lib/x86_64-linux-gnu/librt.so.1 (0x00007f8878928000)
    libstdc++.so.6 => /usr/lib/x86_64-linux-gnu/libstdc++.so.6 (0x00007f887859f000)
    libgcc_s.so.1 => /lib/x86_64-linux-gnu/libgcc_s.so.1 (0x00007f8878387000)
     /lib64/ld-linux-x86-64.so.2 (0x00007f88816a6000)

[Luis]

luis18@luis18:~/cudalucas-code-r102-trunk$ ./CUDALucas 57885161

CUDALucas v2.06beta 64-bit build, compiled Apr  4 2019 @ 13:12:57

binary compiled for CUDA  9.10
CUDA runtime version      9.10
CUDA driver version       10.0

------- DEVICE 0 -------
name                GeForce GTX 1060 3GB
UUID                GPU-d0279b05-7ff6-cb33-d6c8-f7d23090c121
ECC Support?        Disabled
Compatibility       6.1
clockRate (MHz)     1835
memClockRate (MHz)  4004
totalGlobalMem      3165716480
totalConstMem       65536
l2CacheSize         1572864
sharedMemPerBlock   49152
regsPerBlock        65536
warpSize            32
memPitch            2147483647
maxThreadsPerBlock  1024
maxThreadsPerMP     2048
multiProcessorCount 9
maxThreadsDim[3]    1024,1024,64
maxGridSize[3]      2147483647,65535,65535
textureAlignment    512
deviceOverlap       1
pciDeviceID         0
pciBusID            1

You may experience a small delay on 1st startup to due to Just-in-Time Compilation

*** buffer overflow detected ***: ./CUDALucas terminated
Annullato (core dump creato)

luis18@luis18:~/cudalucas-code-r102-trunk$ ./CUDALucas 9689

Idem.

[kriesel]

I think your CUDALucas issue posts would be better placed in the CUDALucas specific thread.

[bayanne]

I thought I would ask again if anyone has got Mersenne in any form running on a AMD GPU on a Mac

[M344587487]

Originally Posted by bayanne

I thought I would ask again if anyone has got Mersenne in any form running on a AMD GPU on a Mac

If you have a new enough OpenCL implementation getting gpuowl to compile should be as simple as editing the Makefile to point to its location. Unfortunately if this link it correct ( https://support.apple.com/en-gb/HT202823 ) Apple's disdain for cross-compatibility has claimed another victim as you need OpenCL 2.0 support for gpuowl ( https://www.mersenneforum.org/showpo...&postcount=930 ).

[bayanne]

Thanks for that, I was contemplating moving to AMD after they did not update NVIDIA drivers in Mojave, seems that they are not taking GPU users very seriously at present ...