[petsc-dev] Kokkos/Crusher perforance
Barry Smith
bsmith at petsc.dev
Sun Jan 23 21:37:19 CST 2022
> On Jan 23, 2022, at 10:01 PM, Junchao Zhang <junchao.zhang at gmail.com> wrote:
>
>
>
> On Sat, Jan 22, 2022 at 9:00 PM Junchao Zhang <junchao.zhang at gmail.com <mailto:junchao.zhang at gmail.com>> wrote:
>
>
>
> On Sat, Jan 22, 2022 at 5:00 PM Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> wrote:
>
> The GPU flop rate (when 100 percent flops on the GPU) should always be higher than the overall flop rate (the previous column). For large problems they should be similar, for small problems the GPU one may be much higher.
>
> If the CPU one is higher (when 100 percent flops on the GPU) something must be wrong with the logging. I looked at the code for the two cases and didn't see anything obvious.
>
> Junchao and Jacob,
> I think some of the timing code in the Kokkos interface is wrong.
>
> * The PetscLogGpuTimeBegin/End should be inside the viewer access code not outside it. (The GPU time is an attempt to best time the kernels, not other processing around the use of the kernels, that other stuff is captured in the general LogEventBegin/End.
> What about potential host to device memory copy before calling a kernel? Should we count it in the kernel time?
Nope, absolutely not. The GPU time represents the time the GPU is doing active work. The outer LogEventBegin/End captures the entire time, including copies, kernel launches etc. No reason to put the copy time in the GPU time because then there would be no need for the GPU since it would be the LogEventBegin/End. The LogEventBegin/End minus the GPU time represents any overhead from transfers.
>
> Good point
> * The use of WaitForKokkos() is confusing and seems inconsistent.
> I need to have a look. Until now, I have not paid much attention to kokkos profiling.
> -For example it is used in VecTDot_SeqKokkos() which I would think has a barrier anyways because it puts a scalar result into update?
> -Plus PetscLogGpuTimeBegin/End is suppose to already have suitable system (that Hong added) to ensure the kernel is complete; reading the manual page and looking at Jacobs cupmcontext.hpp it seems to be there so I don't think WaitForKokkos() is needed in most places (or is Kokkos asynchronous and needs this for correctness?)
> But these won't explain the strange result of overall flop rate being higher than GPU flop rate.
>
> Barry
>
>
>
>
>
>> On Jan 22, 2022, at 11:44 AM, Mark Adams <mfadams at lbl.gov <mailto:mfadams at lbl.gov>> wrote:
>>
>> I am getting some funny timings and I'm trying to figure it out.
>> I figure the gPU flop rates are bit higher because the timers are inside of the CPU timers, but some are a lot bigger or inverted
>>
>> --- Event Stage 2: KSP Solve only
>>
>> MatMult 400 1.0 1.0094e+01 1.2 1.07e+11 1.0 3.7e+05 6.1e+04 0.0e+00 2 55 62 54 0 68 91100100 0 671849 857147 0 0.00e+00 0 0.00e+00 100
>> MatView 2 1.0 4.5257e-03 2.5 0.00e+00 0.0 0.0e+00 0.0e+00 2.0e+00 0 0 0 0 0 0 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> KSPSolve 2 1.0 1.4591e+01 1.1 1.18e+11 1.0 3.7e+05 6.1e+04 1.2e+03 2 60 62 54 60 100100100100100 512399 804048 0 0.00e+00 0 0.00e+00 100
>> SFPack 400 1.0 2.4545e-03 1.4 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> SFUnpack 400 1.0 9.4637e-05 1.7 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> VecTDot 802 1.0 3.0577e+00 2.1 3.36e+09 1.0 0.0e+00 0.0e+00 8.0e+02 0 2 0 0 40 13 3 0 0 67 69996 488328 0 0.00e+00 0 0.00e+00 100
>> VecNorm 402 1.0 1.9597e+00 3.4 1.69e+09 1.0 0.0e+00 0.0e+00 4.0e+02 0 1 0 0 20 6 1 0 0 33 54744 571507 0 0.00e+00 0 0.00e+00 100
>> VecCopy 4 1.0 1.7143e-0228.6 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> VecSet 4 1.0 3.8051e-0316.9 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 0 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> VecAXPY 800 1.0 8.6160e-0113.6 3.36e+09 1.0 0.0e+00 0.0e+00 0.0e+00 0 2 0 0 0 6 3 0 0 0 247787 448304 0 0.00e+00 0 0.00e+00 100
>> VecAYPX 398 1.0 1.6831e+0031.1 1.67e+09 1.0 0.0e+00 0.0e+00 0.0e+00 0 1 0 0 0 5 1 0 0 0 63107 77030 0 0.00e+00 0 0.00e+00 100
>> VecPointwiseMult 402 1.0 3.8729e-01 9.3 8.43e+08 1.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 2 1 0 0 0 138502 262413 0 0.00e+00 0 0.00e+00 100
>> VecScatterBegin 400 1.0 1.1947e+0035.1 0.00e+00 0.0 3.7e+05 6.1e+04 0.0e+00 0 0 62 54 0 5 0100100 0 0 0 0 0.00e+00 0 0.00e+00 0
>> VecScatterEnd 400 1.0 6.2969e+00 8.8 0.00e+00 0.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 10 0 0 0 0 0 0 0 0.00e+00 0 0.00e+00 0
>> PCApply 402 1.0 3.8758e-01 9.3 8.43e+08 1.0 0.0e+00 0.0e+00 0.0e+00 0 0 0 0 0 2 1 0 0 0 138396 262413 0 0.00e+00 0 0.00e+00 100
>> ---------------------------------------------------------------------------------------------------------------------------------------------------------------
>>
>>
>> On Sat, Jan 22, 2022 at 11:10 AM Junchao Zhang <junchao.zhang at gmail.com <mailto:junchao.zhang at gmail.com>> wrote:
>>
>>
>>
>> On Sat, Jan 22, 2022 at 10:04 AM Mark Adams <mfadams at lbl.gov <mailto:mfadams at lbl.gov>> wrote:
>> Logging GPU flops should be inside of PetscLogGpuTimeBegin()/End() right?
>> No, PetscLogGpuTime() does not know the flops of the caller.
>>
>>
>> On Fri, Jan 21, 2022 at 9:47 PM Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> wrote:
>>
>> Mark,
>>
>> Fix the logging before you run more. It will help with seeing the true disparity between the MatMult and the vector ops.
>>
>>
>>> On Jan 21, 2022, at 9:37 PM, Mark Adams <mfadams at lbl.gov <mailto:mfadams at lbl.gov>> wrote:
>>>
>>> Here is one with 2M / GPU. Getting better.
>>>
>>> On Fri, Jan 21, 2022 at 9:17 PM Barry Smith <bsmith at petsc.dev <mailto:bsmith at petsc.dev>> wrote:
>>>
>>> Matt is correct, vectors are way too small.
>>>
>>> BTW: Now would be a good time to run some of the Report I benchmarks on Crusher to get a feel for the kernel launch times and performance on VecOps.
>>>
>>> Also Report 2.
>>>
>>> Barry
>>>
>>>
>>>> On Jan 21, 2022, at 7:58 PM, Matthew Knepley <knepley at gmail.com <mailto:knepley at gmail.com>> wrote:
>>>>
>>>> On Fri, Jan 21, 2022 at 6:41 PM Mark Adams <mfadams at lbl.gov <mailto:mfadams at lbl.gov>> wrote:
>>>> I am looking at performance of a CG/Jacobi solve on a 3D Q2 Laplacian (ex13) on one Crusher node (8 GPUs on 4 GPU sockets, MI250X or is it MI200?).
>>>> This is with a 16M equation problem. GPU-aware MPI and non GPU-aware MPI are similar (mat-vec is a little faster w/o, the total is about the same, call it noise)
>>>>
>>>> I found that MatMult was about 3x faster using 8 cores/GPU, that is all 64 cores on the node, then when using 1 core/GPU. With the same size problem of course.
>>>> I was thinking MatMult should be faster with just one MPI process. Oh well, worry about that later.
>>>>
>>>> The bigger problem, and I have observed this to some extent with the Landau TS/SNES/GPU-solver on the V/A100s, is that the vector operations are expensive or crazy expensive.
>>>> You can see (attached) and the times here that the solve is dominated by not-mat-vec:
>>>>
>>>> ------------------------------------------------------------------------------------------------------------------------
>>>> Event Count Time (sec) Flop --- Global --- --- Stage ---- Total GPU - CpuToGpu - - GpuToCpu - GPU
>>>> Max Ratio Max Ratio Max Ratio Mess AvgLen Reduct %T %F %M %L %R %T %F %M %L %R Mflop/s Mflop/s Count Size Count Size %F
>>>> ---------------------------------------------------------------------------------------------------------------------------------------------------------------
>>>> 17:15 main= /gpfs/alpine/csc314/scratch/adams/petsc/src/snes/tests/data$ grep "MatMult 400" jac_out_00*5_8_gpuawaremp*
>>>> MatMult 400 1.0 1.2507e+00 1.3 1.34e+10 1.1 3.7e+05 1.6e+04 0.0e+00 1 55 62 54 0 27 91100100 0 668874 0 0 0.00e+00 0 0.00e+00 100
>>>> 17:15 main= /gpfs/alpine/csc314/scratch/adams/petsc/src/snes/tests/data$ grep "KSPSolve 2" jac_out_001*_5_8_gpuawaremp*
>>>> KSPSolve 2 1.0 4.4173e+00 1.0 1.48e+10 1.1 3.7e+05 1.6e+04 1.2e+03 4 60 62 54 61 100100100100100 208923 1094405 0 0.00e+00 0 0.00e+00 100
>>>>
>>>> Notes about flop counters here,
>>>> * that MatMult flops are not logged as GPU flops but something is logged nonetheless.
>>>> * The GPU flop rate is 5x the total flop rate in KSPSolve :\
>>>> * I think these nodes have an FP64 peak flop rate of 200 Tflops, so we are at < 1%.
>>>>
>>>> This looks complicated, so just a single remark:
>>>>
>>>> My understanding of the benchmarking of vector ops led by Hannah was that you needed to be much
>>>> bigger than 16M to hit peak. I need to get the tech report, but on 8 GPUs I would think you would be
>>>> at 10% of peak or something right off the bat at these sizes. Barry, is that right?
>>>>
>>>> Thanks,
>>>>
>>>> Matt
>>>>
>>>> Anway, not sure how to proceed but I thought I would share.
>>>> Maybe ask the Kokkos guys if the have looked at Crusher.
>>>>
>>>> Mark
>>>> --
>>>> 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
>>>>
>>>> https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>
>>>
>>> <jac_out_001_kokkos_Crusher_6_8_gpuawarempi.txt>
>>
>> <jac_out_001_kokkos_Crusher_5_8_notpl.txt><jac_out_001_kokkos_Crusher_6_8_notpl.txt>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20220123/2d8c23a6/attachment.html>
More information about the petsc-dev
mailing list