Benchmarks?

[toughengineer]

Benchmarks?

[ColinH]

These are microbenchmarks that we used during early development. They aren't really intended for end-users. That said:

When you build with make the performance test binaries are automatically built, too. If you build with Cmake then ... the easiest way is probably to move them to src/example/json and modify the CMakeLists.txt file accordingly.

Most of the benchmarks need to be called in the shell with a (large) JSON file, like the ones in the tests directery, to do their thing.

[toughengineer]

Do you plan to provide a way to build existing benchmarks with CMake out of the box? Or add new benchmarks?

I have an idea how to make parsing of JSON faster and wanted to experiment. I would like to use common reference point in the form of benchmarks already set up to be built instead of everybody inventing their own way.

[ColinH]

There are no current plans in that direction. Did you manage to build (and run) the benchmarks?

[toughengineer]

I did not try to build it, I guess I will maybe come back to it some other time.

The idea I wanted to experiment with is to parse first small number of JSON array elements (say, 42) into an array (e.g. std::array or plain C-style array, without memory (re)allocation) eventually moving them into the array storage in the value, and only then use push_back/emplace_back. First few push_backs are relatively expensive so temporarily putting parsed elements into an array offsets this cost and makes parsing noticeably faster.

[ColinH]

I just committed an update to the Cmake configuration that also builds the perf applications. If I remember correctly the only reason to exclude them was that the timing code used to be partially platform dependent, but it has since been changed to use std::chrono and should compile everywhere.

[ColinH]

Even assuming that you will be using a bulk insert into the vector, you are still copying the elements around and I wonder whether it wouldn't be more efficient to just call reserve() on it before starting to emplace_back() elements. Should be easy to add to the tao::json::events::set_basic_value::begin_array() Events method; the one that takes a std::size_t already does just that.

[toughengineer]

Thank you, I'll try to come back to it soon.

Calling reserve() is a potential waste of memory — premature optimization.
In the end it's a tradeoff, if you are OK to maybe waste memory in order to maybe gain speed then go for it. I advocate against using reserve() unless you know the exact number of elements you are going to push_back().

[ColinH]

Yes, the trade-off between time and space, as old as computer programming ;-)