My trial of the 1brc challenge in Fortran.
I have been using Fortran occasionally for HPC calculations and thought "why not".
The ideas are:
- mmap the file;
- split the file in chunks and do OpenMP over chunks;
- find position of next ';' and use a hash to get id;
- do branchless scan of temperature
My typical time is around 3 s on 16 threads. Under -O3 the compiler does a lot of optimization and vectorization. While AVX instructions are not used, though. I've tried several tricks to speed-up individual places, in particular tried to convince the compiler to use SIMD for FINDLOC. Did not help at all.
Make sure to rename or make a symlink the input data file (measurements_100k.txt from this repo) as measurements.txt and have it in the cwd.
The major time is taken by findloc(huge_array, semicolon) function (~1.6 s out of 3 s).
There was quite intense discussions of how to improve findloc, but nothing helped.
Tried gfortran, nvfortran, ifx and ifort; gfortran gives much faster (~3 s vs ~12 s) than the others.