Different Results on different Architectures

[N40]

Issue

I find microscopic differences in routing of osrm-extract on different Architectures

Steps to reproduce

run osrm-extract using a Homebrew installation vs a Linux self-compiled one.

Specifications

• v5.7.1
• Mac: 5.27.1_4
• Ubuntu: 5.27.1, self compiled

Questions

1. Is OSRM intended to be machine independent (roundings, rng etc)
2. How could I further debug this?

[DennisOSRM]

What are those macroscopic differences, and what are the architectures involved?

[N40]

Case A: Different Step count

+ Point(10.20715521 53.67772166) -> Point(10.23934191 53.67173931): 411 s , 3353 m, 15 steps
- Point(10.20715521 53.67772166) -> Point(10.23934191 53.67173931): 411 s , 3148 m, 14 steps

Case B: Different Route Distance

+ Point(10.20715521 53.67772166) -> Point(10.23934191 53.67173931): 411 s , 3148 m, 14 steps
- Point(10.20715521 53.67772166) -> Point(10.23934191 53.67173931): 411 s , 3353 m, 15 steps

I also noticed that the following files differ between two OSRM builds

_.osrm.ebg_nodes
_.osrm.edges
_.osrm.fileIndex
_.osrm.icd
_.osrm.tld

Not in the file size, but in the binary file content.

What I am wondering is, is if the exact same OSRM source code and library dependency version (Boost etc) yields deterministic results independent of the architecture it is built on, or if there is some sort of machine specific calculation results, like float implementations or RNGs.