Msieve with GPU support

[jrk]

Jason, when running msieve-gpu, it uses 100% of CPU. Is there a way to make it use less CPU and only do computation on GPU?

[jasonp]

Beyond telling the CUDA driver that the process should not spin waiting for kernel calls to complete, I don't know how to force the application to yield the CPU while the graphics card is busy. Maybe you can run with -p to switch msieve to idle priority? I see 50-80% utilization on a dual core while poly selection is running, with occaisional bursts up to 100%. Your card is a lot nicer than mine, so maybe the CPU has to work harder to keep up :)

[jrk]

Originally Posted by jasonp

Beyond telling the CUDA driver that the process should not spin waiting for kernel calls to complete,

Yes, getting rid of spin-waiting would be ideal. I don't know much of CUDA development, so I don't know how to do that.

Originally Posted by jasonp

I don't know how to force the application to yield the CPU while the graphics card is busy. Maybe you can run with -p to switch msieve to idle priority?

That would still rob time from other idle tasks, so is not ideal.

Originally Posted by jasonp

I see 50-80% utilization on a dual core while poly selection is running, with occaisional bursts up to 100%. Your card is a lot nicer than mine, so maybe the CPU has to work harder to keep up :)

Hmm. Well it is a 8800GTS so I don't think it is better than your 9800GT. So the difference in CPU consumption is for some other reason.

[axn]

commencing number field sieve polynomial selection
time limit set to 300.00 hours
searching leading coefficients from 1 to 2115959
deadline: 1600 seconds per coefficient
coeff 60-960 395459735 435005708 435005709 478506279
------- 395459735-435005708 435005709-478506279
poly  1 p 396019553 q 435057283 coeff 172291190743054499
poly  8 p 395868619 q 435084233 coeff 172236194466384227
poly 14 p 395853127 q 435133843 coeff 172249092415077061
poly 10 p 395661803 q 435139097 coeff 172167919674811891

Seeing output like the above. Does that mean it is currently doing stage1 for the range 60-960 using GPU and will only commence stage 2 after the GPU has done its job?

[ltd]

My GTX260 also uses one complete "virtual core". Its a i7 so it has 4 cores with HT making it 8 "virtual cores".

[frmky]

Originally Posted by jasonp

Beyond telling the CUDA driver that the process should not spin waiting for kernel calls to complete, I don't know how to force the application to yield the CPU while the graphics card is busy.

In the cuCtxCreate() call in gnfs/poly/stage1/stage1_sieve.c, change CU_CTX_SCHED_YIELD to CU_CTX_BLOCKING_SYNC. This takes the CPU usage from near 100% to almost 0.

As a crude comparison of relative speeds, I created a cpu version by changing the #if statement. The CPU version output 5 "poly" lines in the time that the GPU version output 66 "poly" lines. By this crude measure, it's somewhere around 11-17 times faster.

The S1070 here has 4 GPUs. Specifying the number of GPUs, and perhaps even which GPUs, would be useful.

[jrk]

Originally Posted by frmky

In the cuCtxCreate() call in gnfs/poly/stage1/stage1_sieve.c, change CU_CTX_SCHED_YIELD to CU_CTX_BLOCKING_SYNC. This takes the CPU usage from near 100% to almost 0.

That brings my CPU usage down to 88%.

[xilman]

The good news is that with the assistance of Jason and Greg Childers I managed to compile a binary.

The bad news is:

[pcl@maat msieve-gpu]$ ./msieve -np -i test.num
deadline: 1600 seconds per coefficient
error (line 137): CUDA_ERROR_FILE_NOT_FOUND
[pcl@maat msieve-gpu]$ grep CUDA_ERROR_FILE_NOT_FOUND */*
common/cuda_xface.c:    case CUDA_ERROR_FILE_NOT_FOUND: return "CUDA_ERROR_FILE_NOT_FOUND";
Binary file common/cuda_xface.o matches

Hmm. I'll investigate more tomorrow when I'm likely to be more sober and less tired than I am right now.


Paul

[jasonp]

axn: the code begins stage 2 after every output line appears there; this isn't like pol5, stage 1 output is not batched up and sent to stage 2 all at once.

Greg: thanks, that works much better. Am I correct that I can't use that flag if I want asynchronous behavior, i.e. if I build double-buffering into the code so that the CPU runs stage 2 while the GPU runs the next stage 1?

jrk: my CPU usage drops to 10-15% with Greg's change

Paul: once nvcc generates it, stage1_core.ptx is expected to be in the same directory as the msieve binary

all: binary has been updated again.

[jrk]

Originally Posted by jasonp

Nvidia's CUDA profiler shows the GPU is only 15% utilized, so there's headroom for much more performance too. I'm still working on this, and will post updated binaries as I understand this hardware better.

Just a quick thought: How much time do the kernels spend on the GPU relative to the overhead of loading them?

[jrk]

Originally Posted by xilman

The good news is that with the assistance of Jason and Greg Childers I managed to compile a binary.

The bad news is:

[pcl@maat msieve-gpu]$ ./msieve -np -i test.num
deadline: 1600 seconds per coefficient
error (line 137): CUDA_ERROR_FILE_NOT_FOUND
[pcl@maat msieve-gpu]$ grep CUDA_ERROR_FILE_NOT_FOUND */*
common/cuda_xface.c:    case CUDA_ERROR_FILE_NOT_FOUND: return "CUDA_ERROR_FILE_NOT_FOUND";
Binary file common/cuda_xface.o matches

Hmm. I'll investigate more tomorrow when I'm likely to be more sober and less tired than I am right now.


Paul

You need the stage1_core.ptx file in your msieve directory. If you are on Linux you'll need to compile this file per frmky's instructions in the second post here.