diff --git a/test/perf/README.md b/test/perf/README.md
new file mode 100644
index 00000000..874149c8
--- /dev/null
+++ b/test/perf/README.md
@@ -0,0 +1,58 @@
+# Benchmarks
+
+The operating systems typically caches access to the file system.
+To make these benchmarks more realistic, the file system caches is dropped at every iteration so that the disk IO *is* included in the reported run times.
+On Linux, the caches are dropped by writing `3` to the file `/proc/sys/vm/drop_caches` however this requires super user privileges.
+These benchmarks require a Linux operating system (as dropping file caches is OS-specific).
+
+
+## Installation
+
+### Julia packages
+
+Within a Julia shell install `BenchmarkTools` and `NCDatasets` using these julia commands:
+
+```julia
+using Pkg
+Pkg.add(["BenchmarkTools","NCDatasets"])
+```
+
+### Python packages
+
+Install the python packages `netCDF4` and `numpy` using this shell command:
+
+```bash
+pip install netCDF4 numpy
+```
+
+### R packages
+
+Within a R shell install `microbenchmark` and `ncdf4` using these R commands:
+```R
+install.packages("microbenchmark")
+install.packages("ncdf4")
+```
+
+## Running the benchmark
+
+These are the steps to run the benchmark:
+
+* Prepare the file `filename_fv.nc` with:
+
+```bash
+julia generate_data.jl
+```
+
+* As a *root user*, run the shell script `benchmark.sh`. It is necessary that the root user has access to the Julia, python and R netCDF packages (NCDatasets, netCDF4 and ncdf4 respectively).
+
+```bash
+./benchmark.sh
+```
+
+If all packages are installed in the home directory of an unpriviledges user e.g. `my_user_name`, they can be made available to the root user changing temporarily the `HOME` environement variable to `/home/my_user_name` in the root shell before running `./benchmark.sh`:
+
+```bash
+HOME=/home/my_user_name ./benchmark.sh
+```
+
+The script will output a markdown table with the benchmark statistics.