1BRC in Fortran (~3 s on 16 threads)

[sshestov]

https://github.com/sshestov/1brc-fortran

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 the next ';' via findloc;
• calculate id-hash;
• 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.

I don't do final sorting though.