search

madgraph5 / madgraph4gpu

GPU development for the Madgraph5_aMC@NLO event generator software package
30 stars 33 forks source link

Cuda time profiles for DY+3j have high non-ME component #994

Open valassi opened 2 weeks ago

valassi commented 2 weeks ago

Yesterday I ran some very first tests of cuda DY+3j with (OLD) timers in PR #948.

The cuda profiles are clearly weird

This is for 500 events

[avalassi@itscrd90 bash] /data/avalassi/GPU2023/madgraph4gpuX/epochX/cudacpp/tlau/fromgridpacks> more pp_dy3j.mad/summary.txt 
pp_dy3j.mad/fortran/output.txt
[GridPackCmd.launch] OVERALL TOTAL    1945.6279 seconds
[madevent COUNTERS]  PROGRAM TOTAL    1910.3
[madevent COUNTERS]  Fortran Overhead 665.412
[madevent COUNTERS]  Fortran MEs      1244.89
--------------------------------------------------------------------------------
pp_dy3j.mad/cppnone/output.txt
[GridPackCmd.launch] OVERALL TOTAL    1920.0969 seconds
[madevent COUNTERS]  PROGRAM TOTAL    1896.82
[madevent COUNTERS]  Fortran Overhead 668.916
[madevent COUNTERS]  CudaCpp MEs      1223.65
[madevent COUNTERS]  CudaCpp HEL      4.2527
--------------------------------------------------------------------------------
pp_dy3j.mad/cppsse4/output.txt
[GridPackCmd.launch] OVERALL TOTAL    1336.0181 seconds
[madevent COUNTERS]  PROGRAM TOTAL    1313.34
[madevent COUNTERS]  Fortran Overhead 668.988
[madevent COUNTERS]  CudaCpp MEs      642.063
[madevent COUNTERS]  CudaCpp HEL      2.2873
--------------------------------------------------------------------------------
pp_dy3j.mad/cppavx2/output.txt
[GridPackCmd.launch] OVERALL TOTAL    960.2111 seconds
[madevent COUNTERS]  PROGRAM TOTAL    937.127
[madevent COUNTERS]  Fortran Overhead 667.996
[madevent COUNTERS]  CudaCpp MEs      267.903
[madevent COUNTERS]  CudaCpp HEL      1.2269
--------------------------------------------------------------------------------
pp_dy3j.mad/cpp512y/output.txt
[GridPackCmd.launch] OVERALL TOTAL    940.0347 seconds
[madevent COUNTERS]  PROGRAM TOTAL    917.336
[madevent COUNTERS]  Fortran Overhead 668.996
[madevent COUNTERS]  CudaCpp MEs      247.179
[madevent COUNTERS]  CudaCpp HEL      1.1605
--------------------------------------------------------------------------------
pp_dy3j.mad/cpp512z/output.txt
[GridPackCmd.launch] OVERALL TOTAL    1022.0703 seconds
[madevent COUNTERS]  PROGRAM TOTAL    997.125
[madevent COUNTERS]  Fortran Overhead 669.147
[madevent COUNTERS]  CudaCpp MEs      326.476
[madevent COUNTERS]  CudaCpp HEL      1.503
--------------------------------------------------------------------------------
pp_dy3j.mad/cuda/output.txt
[GridPackCmd.launch] OVERALL TOTAL    969.4855 seconds
[madevent COUNTERS]  PROGRAM TOTAL    853.823
[madevent COUNTERS]  Fortran Overhead 826.381
[madevent COUNTERS]  CudaCpp MEs      7.865
[madevent COUNTERS]  CudaCpp HEL      19.578
--------------------------------------------------------------------------------
valassi commented 2 weeks ago
  • there is a high non-ME component (here stil called 'fortran overhead', these are olf timers)

specifically, fortran and cpp have 668s, cuda has 826

  • there is a high outside-madevent ('python/bash'? time spent deleting the applications??) component

specifically, fortran has 1945-1910 i.e 35s, cuda has 969-853 i.e. 116s

valassi commented 1 week ago

I have stripped off the python/bash component to #1000 (for cuda but not only!). Instead here I keep only the non-ME madevent component (in cuda).