Logging GPU flops should be inside of PetscLogGpuTimeBegin()/End() right? On Fri, Jan 21, 2022 at 9:47 PM Barry Smith <bsm...@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 <mfad...@lbl.gov> wrote: > > Here is one with 2M / GPU. Getting better. > > On Fri, Jan 21, 2022 at 9:17 PM Barry Smith <bsm...@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 <knep...@gmail.com> wrote: >> >> On Fri, Jan 21, 2022 at 6:41 PM Mark Adams <mfad...@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> > > >
