Why unit of weight/duration is 100ms?

[wangyoucao577]

When I read code, I found that the unit of weight/duration seems 100ms instead of seconds.

I found below comments from code:

osrm-backend/include/util/typedefs.hpp

Line 81 in e86d937

using TurnPenalty = std::int16_t; // turn penalty in 100ms units

using TurnPenalty = std::int16_t; // turn penalty in 100ms units

Also, I found a lot of hardcoded codes multiply weight/duration by 10, e.g.

osrm-backend/src/extractor/graph_compressor.cpp

Line 253 in e86d937

node_duration_penalty = extraction_turn.duration * 10;

                    node_duration_penalty = extraction_turn.duration * 10;
                    node_weight_penalty = extraction_turn.weight * weight_multiplier;

osrm-backend/src/extractor/extraction_containers.cpp

Line 397 in a1e5061

edge.weight = std::max<EdgeWeight>(1, std::round(segment.weight * weight_multiplier));

            edge.weight = std::max<EdgeWeight>(1, std::round(segment.weight * weight_multiplier));
            edge.duration = std::max<EdgeWeight>(1, std::round(segment.duration * 10.));

I think use the 100ms unit is a little strange. Why did this?

Moreover, I wrote down a document How OSRM Calculate Weight and Duration when I reading codes. FYI. Please let me know if anything not correct, thanks!

[danpat]

@wangyoucao577 This is done because OSRM uses fixed point arithmetic for duration and weight calculations.

The decision to do it this way pre-dates my participation in the project, but if I had to guess, I'd say it was a performance choice - historically, integer math was faster than floating point math on most CPUs (although the reality of that on modern hardware is complicated).

I don't know that any deep justification for this choice has ever been given. It would certainly be possible to use a float type instead, but it would likely come with a myriad of subtle changes to behaviour, due to floating point precision behaviour. I cannot say whether the changes would be significant or not. 🤷‍♂️

[wangyoucao577]

Thanks @danpat for the quick reply. It makes sense due to historically chosen. Agree with no need to change to float by now.
For the multiply 10 part, especially for duration * 10, I think it'll be better to use macro/simple function instead. There're too many hardcodes in repo by now. I'll try to create PR when I have time.