[petsc-dev] [petsc-maint #88993] Petsc with Cuda 4.0 and Multiple GPUs

[Barry Smith]

On Oct 2, 2011, at 6:39 PM, Dave Nystrom wrote:

> Thanks for the update.  I don't believe I have gotten a run with good
> performance yet, either from C or Fortran.  I wish there was an easy way for
> me to force use of only one of my gpus.  I don't want to have to pull one of
> the gpus in order to see if that is complicating things with Cuda 4.0.  I'll
> be eager to hear if you make any progress on figuring things out.
> 
> Do you understand yet why the petsc ex2.c example is able to parse the
> "-cuda_show_devices" argument but ex2f.F does not?

   Matt put the code in the wrong place in PETSc, that is all, no big existentialist reason. I will fix that.


   Barry

> 
> Thanks,
> 
> Dave
> 
> Barry Smith writes:
>> It is not doing the MatMult operation on the GPU and hence needs to move
>> the vectors back and forth for each operation (since MatMult is done on
>> the CPU with the vector while vector operations are done on the GPU) hence
>> the terrible performance.
>> 
>> Not sure why yet. It is copying the Mat down for me from C.
>> 
>> Barry
>> 
>> On Oct 2, 2011, at 4:18 PM, Dave Nystrom wrote:
>> 
>>> In case it might be useful, I have attached two log files of runs with the
>>> ex2f petsc example from src/ksp/ksp/examples/tutorials.  One was run back in
>>> April with petsc-dev linked to Cuda 3.2.  It shows excellent runtime
>>> performance.  The other was run today with petsc-dev checked out of the
>>> mercurial repo yesterday morning and linked to Cuda 4.0.  In addition to the
>>> differences in run time performance, I also do not see an entry for
>>> MatCUSPCopyTo in the profiling section.  I'm not sure what the significance
>>> of that is.  I do observe that the run time for PCApply is about the same for
>>> the two cases.  I think I would expect that to be the case even if the
>>> problem were partitioned across two gpus.  However, it does make me wonder if
>>> the absence of MatCUSPCopyTo in the profiling section of the Cuda 4.0 log
>>> file is an indication that the matrix was not actually copied to the gpu.
>>> I'm not sure yet how to check for that.  Hope this might be useful.
>>> 
>>> Thanks,
>>> 
>>> Dave
>>> 
>>> 
>>> <ex2f_3200_3200_cuda_yes_cuda_3.2.log><ex2f_3200_3200_cuda_yes_cuda_4.0.log>
>>> Dave Nystrom writes:
>>>> Matthew Knepley writes:
>>>>> On Sat, Oct 1, 2011 at 11:26 PM, Dave Nystrom <Dave.Nystrom at tachyonlogic.com> wrote:
>>>>>> Barry Smith writes:
>>>>>>> On Oct 1, 2011, at 9:22 PM, Dave Nystrom wrote:
>>>>>>>> Hi Barry,
>>>>>>>> 
>>>>>>>> I've sent a couple more emails on this topic.  What I am trying to do at the
>>>>>>>> moment is to figure out how to have a problem run on only one gpu if it will
>>>>>>>> fit in the memory of that gpu.  Back in April when I had built petsc-dev with
>>>>>>>> Cuda 3.2, petsc would only use one gpu if you had multiple gpus on your
>>>>>>>> machine.  In order to use multiple gpus for a problem, one had to use
>>>>>>>> multiple threads with a separate thread assigned to control each gpu.  But
>>>>>>>> Cuda 4.0 has, I believe, made that transparent and under the hood.  So now
>>>>>>>> when I run a small example problem such as
>>>>>>>> src/ksp/ksp/examples/tutorials/ex2f.F with an 800x800 problem, it gets
>>>>>>>> partitioned to run on both of the gpus in my machine.  The result is a very
>>>>>>>> large performance hit because of communication back and forth from one gpu to
>>>>>>>> the other via the cpu.
>>>>>>> 
>>>>>>> How do you know there is lots of communication from the GPU to the CPU? In
>>>>>>> the -log_summary? Nope because PETSc does not manage anything like that
>>>>>>> (that is one CPU process using both GPUs).
>>>>>> 
>>>>>> What I believe is that it is being managed by Cuda 4.0, not by petsc.
>>>>>> 
>>>>>>>> So this problem with a 3200x3200 grid runs 5x slower
>>>>>>>> now than it did with Cuda 3.2.  I believe if one is programming down at the
>>>>>>>> cuda level, it is possible to have a smaller problem run on only one gpu so
>>>>>>>> that there is communication only between the cpu and gpu and only at the
>>>>>>>> start and end of the calculation.
>>>>>>>> 
>>>>>>>> To me, it seems like what is needed is a petsc option to specify the number
>>>>>>>> of gpus to run on that can somehow get passed down to the cuda level through
>>>>>>>> cusp and thrust.  I fear that the short term solution is going to have to be
>>>>>>>> for me to pull one of the gpus out of my desktop system but it would be nice
>>>>>>>> if there was a way to tell petsc and friends to just use one gpu when I want
>>>>>>>> it to.
>>>>>>>> 
>>>>>>>> If necessary, I can send a couple of log files to demonstrate what I am
>>>>>>>> trying to describe regarding the performance hit.
>>>>>>> 
>>>>>>> I am not convinced that the poor performance you are getting now has
>>>>>>> anything to do with using both GPUs. Please run a PETSc program with the
>>>>>>> command -cuda_show_devices
>>>>>> 
>>>>>> I ran the following command:
>>>>>> 
>>>>>> ex2f -m 8 -n 8 -ksp_type cg -pc_type jacobi -log_summary -cuda_show_devices
>>>>>> -mat_type aijcusp -vec_type cusp -options_left
>>>>>> 
>>>>>> The result was a report that there was one option left, that being
>>>>>> -cuda_show_devices.  I am using a copy of petsc-dev that I cloned and built
>>>>>> this morning.
>>>>> 
>>>>> What do you have at src/sys/object/pinit.c:825? You should see the code
>>>>> that processes this option. You should be able to break there in the
>>>>> debugger and see what happens. This sounds again like you are not
>>>>> processing options correctly.
>>>> 
>>>> Hi Matt,
>>>> 
>>>> I'll take a look at that in a bit and see if I can figure out what is going
>>>> on.  I do see the code that you mention that processes the arguments that
>>>> Barry mentioned.  In terms of processing options correctly, at least in this
>>>> case I am actually running one of the petsc examples rather than my own
>>>> code.  And it seems to correctly process the other command line arguments.
>>>> Anyway, I'll write more after I have had a chance to investigate more.
>>>> 
>>>> Thanks,
>>>> 
>>>> Dave
>>>> 
>>>>> Matt
>>>>> 
>>>>>>> What are the choices?  You can then pick one of them and run with
>>>>>> -cuda_set_device integer
>>>>>> 
>>>>>> The -cuda_set_device option does not appear to be recognized either, even
>>>>>> if I choose an integer like 0.
>>>>>> 
>>>>>>> Does this change things?
>>>>>> 
>>>>>> I suspect it would change things if I could get it to work.
>>>>>> 
>>>>>> Thanks,
>>>>>> 
>>>>>> Dave
>>>>>> 
>>>>>>> Barry
>>>>>>> 
>>>>>>>> 
>>>>>>>> Thanks,
>>>>>>>> 
>>>>>>>> Dave
>>>>>>>> 
>>>>>>>> Barry Smith writes:
>>>>>>>>> Dave,
>>>>>>>>> 
>>>>>>>>> We have no mechanism in the PETSc code for a PETSc single CPU process to
>>>>>>>>> use two GPUs at the same time. However you could have two MPI processes
>>>>>>>>> each using their own GPU.
>>>>>>>>> 
>>>>>>>>> The one tricky part is you need to make sure each MPI process uses a
>>>>>>>>> different GPU. We currently do not have a mechanism to do this assignment
>>>>>>>>> automatically. I think it can be done with cudaSetDevice(). But I don't
>>>>>>>>> know the details, sending this to petsc-dev at mcs.anl.gov where more people
>>>>>>>>> may know.
>>>>>>>>> 
>>>>>>>>> PETSc-folks,
>>>>>>>>> 
>>>>>>>>> We need a way to have this setup automatically.
>>>>>>>>> 
>>>>>>>>> Barry
>>>>>>>>> 
>>>>>>>>> On Oct 1, 2011, at 5:43 PM, Dave Nystrom wrote:
>>>>>>>>> 
>>>>>>>>>> I'm running petsc on a machine with Cuda 4.0 and 2 gpus.  This is a desktop
>>>>>>>>>> machine with a single processor.  I know that Cuda 4.0 has support for
>>>>>>>>>> running on multiple gpus but don't know if petsc uses that.  But suppose I
>>>>>>>>>> have a problem that will fit in the memory for a single gpu.  Will petsc run
>>>>>>>>>> the problem on a single gpu or does it split it between the 2 gpus and incur
>>>>>>>>>> the communication overhead of copying data between the two gpus?
>>>>>>>>>> 
>>>>>>>>>> Thanks,
>>>>>>>>>> 
>>>>>>>>>> Dave
>>>>> 
>>>>> -- 
>>>>> What most experimenters take for granted before they begin their experiments
>>>>> is infinitely more interesting than any results to which their experiments
>>>>> lead.
>>>>> -- Norbert Wiener
>>