Benchmarks
The benchmarks are the result of 1 million iterations per test, with randomly generated tags and values following a gaussian distribution. Thy're are performed on the static library compiled following the installation guide and they measure execution speed in terms of clock cycles rather than seconds, draw your own conclusions according to your CPU clock-speed.
Comments:
- The benchmarks were run using only aligned memory, mostly static, with exception for the actual random strings.
- Serialization takes much more time than deserialization, as it involves copying the data to a buffer.
- Deserialization is generally much faster as it uses zero-copy techniques.
- Direct zero-copy serialization with vectorized writev and no memcpy was attempted but resulted in a 3x performance decrease, likely due to the small nature of the FIX fields and tags.
Deserialization
Serialization
Run your own benchmarks
To run your own benchmarks you can follow the steps below:
- Compile the library as described in the installation guide
- Compile the benchmark target:
cmake --build . --target benchmark - Run the benchmark executable:
./benchmark - Generate a plot with the results:
python3 ./benchmarks/plot.py *.csv