Originally, these were detailed experiments done over the simdjson code base for the paper Parsing gigabytes of JSON per second.
We use a simple Makefile to execute the docker scripts. If you have a proper configured Linux system, you should be able to generate all experimental results with a simple command:
make
The code for the experiments is in the experiments directory. Each experiment is in its own repository with a docker file describing the container.
The experimental results are in the result directory. They are organized first by node name (to distinguish results from different systems). When you type make, it creates a new subdirectory and fills it with experimental results.
There is a distinct directory called analysis that contains the scripts processing the resutlts from the result directory. The analysis scripts have dependencies: python, pandas, gnuplot. Doing the analysis is not fully automated, by design: a human being needs to look at the number.
Software:
- Linux
- Docker (you need privileged access).
Hardware:
- AVX2-compatible processing (Intel, AMD).
- We check that hyperthreading is disabled.
You need to install docker, make sure it is running, and make yourself a priviliged user which may involve usermod -G -a docker myusername or some similar command. You want to create a docker group, put yourself in the docker group and restarts docker.