mfaktc: a CUDA program for Mersenne prefactoring

[Rodrigo]

Mark and James,

Thanks for the ideas.

The PC in question is double-booting Windows XP and Vista. A couple of months ago, I bought a 1050 for another Vista system that needed a new card, but the card wouldn't run on it; there seem to be no Vista drivers for the 1050. So I installed the 1050 in my new Linux box, where it's putting out some 250 GHz-d/day, compared to the 630's 53 GHz-d/day.

I guess I could put Linux on the XP/Vista machine to run another GTX 1050.

P.S. James, that's a pretty neat chart you put together. I didn't know one could do that on your page!

[kriesel]

Originally Posted by Rodrigo

Mark and James,

I guess I could put Linux on the XP/Vista machine to run another GTX 1050.

Or maybe:
linux host OS + GTX1050; mprime and gpu apps run on host OS continuously;
Virtualbox environment;
Guest OS XP; Guest OS Vista; run either or both or neither
(down or pause the guest OSes when not in use, for max GIMPS throughput when system is not in use; preferably down, to release memory for P-1 or ECM if running those)

I've been meaning to play with Virtualbox as a means of readily migrating a complex main PC environment from old hardware set to new.

[Rodrigo]

Originally Posted by kriesel

Or maybe:
linux host OS + GTX1050; mprime and gpu apps run on host OS continuously;
Virtualbox environment;
Guest OS XP; Guest OS Vista; run either or both or neither
(down or pause the guest OSes when not in use, for max GIMPS throughput when system is not in use; preferably down, to release memory for P-1 or ECM if running those)

I've been meaning to play with Virtualbox as a means of readily migrating a complex main PC environment from old hardware set to new.

That's an intriguing idea. I'm still transitioning my work to the Linux computer but when that process is done, I just might look into virtualization. One step at a time...

Wonder how one would migrate an existing hardware PC environment to a virtualized one. Can you take an image of the existing HDD and use it as the basis for the virtual PC?

[Sake]

mfactc compiled with CUDA 9.1 (tested on fedora 27 Linux) doesn't seem to work. Same issues with pre-compiled mfactc versions.
The self-test fails with all kernels except GPU kernel "71bit_mul24". Compilation with different GPU architecture settings (5, 6, etc.) caused the same issues. Tested with CUDA 9.1.85.1, fedora 27 on NVIDIA TITAN V

Is it a CUDA bug? Any idea? Thanks.


# Compiler settings for .cu files (CPU/GPU)
NVCC = nvcc
NVCCFLAGS = $(CUDA_INCLUDE) --ptxas-options=-v -arch=sm_70 -gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_70,code=compute_70


-------------------------------------------------------------------------------------
mfaktc v0.21 (64bit built)

Compiletime options
THREADS_PER_BLOCK 1024
SIEVE_SIZE_LIMIT 32kiB
SIEVE_SIZE 193154bits
SIEVE_SPLIT 250
MORE_CLASSES enabled

Runtime options
SievePrimes 25000
SievePrimesAdjust 1
SievePrimesMin 5000
SievePrimesMax 100000
NumStreams 3
CPUStreams 3
GridSize 3
GPU Sieving enabled
GPUSievePrimes 82486
GPUSieveSize 64Mi bits
GPUSieveProcessSize 16Ki bits
Checkpoints enabled
CheckpointDelay 30s
WorkFileAddDelay 600s
Stages enabled
StopAfterFactor bitlevel
PrintMode full
V5UserID (none)
ComputerID (none)
AllowSleep no
TimeStampInResults no

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

CUDA device info
name Graphics Device
compute capability 7.0
max threads per block 1024
max shared memory per MP 98304 byte
number of multiprocessors 80
clock rate (CUDA cores) 1455MHz
memory clock rate: 850MHz
memory bus width: 3072 bit

Automatic parameters
threads per grid 655360
GPUSievePrimes (adjusted) 82486
GPUsieve minimum exponent 1055144

########## testcase 1/2867 ##########

[chalsall]

Originally Posted by Sake

Is it a CUDA bug? Any idea?

Oliver (the author of mfaktc) confirms that this is a bug in CUDA 9.0.

He has alerted nVidia, and they've acknowledged the problem. Still no working patch yet.

Your only option is to downgrade your CUDA to 8.x (I believe; could be lower) and recompile.

[TheJudger]

@Chris: No, he can't... his Volta GPU needs CUDA 9.0 or newer.

@Sake Above is not that bad, just run CUDALucas on your toy.

Oliver

[kriesel]

Originally Posted by TheJudger

@Chris: No, he can't... his Volta GPU needs CUDA 9.0 or newer.

@Sake Above is not that bad, just run CUDALucas on your toy.

Oliver

Or run cudapm1. It will generate results more frequently than CUDALucas. Not many are running P-1, so it's easy to be in the top 500 producers list. https://www.mersenne.org/report_top_500_p-1/
Manual reporting only though for P-1 on GPU.

[Sake]

Thank you for the help! Will test with cudapm1 in the meanwhile.

[tServo]

Originally Posted by Sake

Thank you for the help! Will test with cudapm1 in the meanwhile.

Sake, have you tried Cuda Toolkit 9.1.85 ?
Your output above from the bug shows 9.10.

[TheJudger]

still broken in 9.1 (Volta only)

Oliver

[ramgeis]

Originally Posted by TheJudger

still broken in 9.1 (Volta only)

Oliver

9.2 is out. Anyone tried to check it already?