m3g / 2021_fortrancon Goto Github PK

I was surprised with the results in the benchmark of Fortran vs Julia (I was expecting similar performance but with a little advantage of Fortran over Julia), so I've decided to rerun the benchmarks with different optimization levels:

gfortran -O0 -march=native particle_simulation.f90 -o particle_simulation
echo "Fortran -O0:"
time ./particle_simulation

gfortran -O1 -march=native particle_simulation.f90 -o particle_simulation
echo "Fortran -O1:"
time ./particle_simulation

gfortran -O2 -march=native particle_simulation.f90 -o particle_simulation
echo "Fortran -O2:"
time ./particle_simulation

gfortran -O3 -march=native particle_simulation.f90 -o particle_simulation
echo "Fortran -O3:"
time ./particle_simulation

gfortran -Ofast -march=native particle_simulation.f90 -o particle_simulation
echo "Fortran -Ofast:"
time ./particle_simulation

echo "Julia:"
time julia particle_simulation.jl

With the results:

Fortran -O0:
real	1m10,165s
user	1m10,128s
sys	0m0,003s

Fortran -O1:
real	0m40,524s
user	0m40,497s
sys	0m0,000s

Fortran -O2:
real	0m36,427s
user	0m36,383s
sys	0m0,017s

Fortran -O3:
real	0m39,488s
user	0m39,443s
sys	0m0,007s

Fortran -Ofast:
real	0m36,217s
user	0m36,184s
sys	0m0,007s

Julia:
real	0m38,021s
user	0m37,924s
sys	0m0,073s

While I know that the intent of this simple benchmark is more something in the order of "Julia speed is comparable to Fortran", I think that the results can mislead the reader into thinking that Julia is faster than Fortran. (The difference is marginal anyways, so if the user is more productive with Julia I wouldn't say that a switch is necessary just based on speed)

Honestly I don't know exactly why, but the "best" optimization algorithm (-O3) ain't necessarily the best, as it can be shown with this implementation. This also shows that there is an extra step to take into account with compiled languages, and that's that compiling flags can make a huge impact in the result, giving another layer of complexity of Fortran over Julia.