fsbench

A set of R-centric benchmarks that evaluate filesystem performance for a number of common tasks. Currently for Linux and macOS only.

Preparation

make setup

This will install required R packages, compile a small binary executable, and download some large data files to run the benchmarks against.

Running make setup will execute a command under sudo. This is necessary to create the tools/purge executable, which is invoked before each benchmark. The purge executable dumps the operating system's disk cache, which is a privileged operation, so the setuid flag to be set.

Running

make

This will run for several minutes and then dump some timing information to the screen.

Run Configuration

Two environment variables can be used to configure fsbench:

• TARGET_DIR defaults to ../fsbench.work and controls where data will be written to/read from. This should be a directory path located on the filesystem you want to test, and needs to be set for both make setup and make. fsbench will attempt to create the directory if it does not exist. This directory must NOT be a subdirectory of the fsbench directory, otherwise the package installation benchmarks will throw confusing errors.

• OUTPUT_FILE defaults to ./results-<date>-<time>.csv and indicates the path where the benchmark results should be recorded, as CSV data. (The same results are always printed to the screen, in tabular form.)

Comparing results

Once you have one or more results files from running the benchmarks, you can use the comparison tool to compare multiple runs across several filesystems. The comparison tool will average multiple runs for each distinct filesystem tested, and produce bar charts to compare performance between the different filesystems.

To run the comparison tool:

Rscript compare.R <filesystem name>=<results file glob> ...

For example, to compare several runs against an SSD to several runs against an EFS volume:

Rscript compare.R ssd=*ssd.csv efs=*efs.csv

This will cause all *ssd.csv files containing ssd results to be loaded and averaged together and graphed against all of the (averaged) efs runs within the *efs.csv files. After the comparison has run, two image files will be generated:

• serial-plot-results.png shows all of the serial tests that have no parallelization, one plot for each test benchmark
• parallel-plot-results.png shows all of the parallel tests, one plot for each test benchmark

License

fsbench is public domain (CC0).