diff --git a/test/perf/README.md b/test/perf/README.md
index 874149c8..f0c59527 100644
--- a/test/perf/README.md
+++ b/test/perf/README.md
@@ -1,11 +1,5 @@
 # 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
@@ -43,16 +37,31 @@ These are the steps to run the benchmark:
 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).
+* Run the shell script `benchmark.sh`.
 
 ```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`:
+The script will output a markdown table with the benchmark statistics.
+
+## Dropping file caches
+
+The operating systems typically caches access to the file system.
+To make these benchmarks more realistic, the file system caches can be dropped at every iteration using the benchmark script with the option `--drop-caches`.
+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.
+In this case, these benchmarks require a Linux operating system (as dropping file caches is OS-specific).
+
+
+* 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
-HOME=/home/my_user_name ./benchmark.sh
+./benchmark.sh --drop-caches
 ```
 
-The script will output a markdown table with the benchmark statistics.
+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 --drop-caches
+```
diff --git a/test/perf/benchmark-R-ncdf4.R b/test/perf/benchmark-R-ncdf4.R
index 584cc3cb..b57a2268 100644
--- a/test/perf/benchmark-R-ncdf4.R
+++ b/test/perf/benchmark-R-ncdf4.R
@@ -6,15 +6,18 @@
 library(ncdf4)
 library(microbenchmark)
 
+print(R.version.string)
 print(paste("ncdf4 version: ",packageVersion("ncdf4")))
 
 fname = "filename_fv.nc"
 
-process <- function(fname) {
-  # drop file caches; requires root
-  fileConn<-file("/proc/sys/vm/drop_caches",open = "wt")
-  writeLines("3", fileConn)
-  close(fileConn)
+process <- function(fname,drop_caches) {
+  if (drop_caches) {
+    # drop file caches; requires root
+    fileConn<-file("/proc/sys/vm/drop_caches",open = "wt")
+    writeLines("3", fileConn)
+    close(fileConn)
+  }
 
   nc = nc_open(fname)
 
@@ -29,14 +32,16 @@ process <- function(fname) {
   return(tot/nmax)
 }
 
+drop_caches <- "--drop-caches" %in% commandArgs(trailingOnly=TRUE)
+print(paste("drop caches: ",drop_caches))
 start_time <- Sys.time()
-tot = process(fname)
+tot = process(fname,drop_caches)
 end_time <- Sys.time()
 print(paste("time ",end_time - start_time))
 
 print(paste("result ",tot))
 
-mbm <- microbenchmark("ncdf4" = process(fname),times=100)
+mbm <- microbenchmark("ncdf4" = process(fname,drop_caches),times=100)
 
 fileConn<-file("R-ncdf4.txt",open = "wt")
 
diff --git a/test/perf/benchmark-julia-NCDatasets.jl b/test/perf/benchmark-julia-NCDatasets.jl
index 347a0505..a1fe4db6 100644
--- a/test/perf/benchmark-julia-NCDatasets.jl
+++ b/test/perf/benchmark-julia-NCDatasets.jl
@@ -13,9 +13,11 @@ function compute(v)
     return tot/size(v,3)
 end
 
-function process(fname)
-    # drop file caches; requires root
-    write("/proc/sys/vm/drop_caches","3")
+function process(fname,drop_caches)
+    if drop_caches
+        # drop file caches; requires root
+        write("/proc/sys/vm/drop_caches","3")
+    end
 
     ds = NCDataset(fname,"r") do ds
         v = ds["v1"];
@@ -24,12 +26,15 @@ function process(fname)
     end
 end
 
-fname = "filename_fv.nc"
-tot = process(fname)
+drop_caches = "--drop-caches" in ARGS
+println("Julia ",VERSION)
+println("drop caches: ",drop_caches)
 
+fname = "filename_fv.nc"
+tot = process(fname,drop_caches)
 println("result ",tot)
 
-bm = run(@benchmarkable process(fname) samples=100 seconds=10000)
+bm = run(@benchmarkable process(fname,drop_caches) samples=100 seconds=10000)
 
 @show bm
 
diff --git a/test/perf/benchmark-python-netCDF4.py b/test/perf/benchmark-python-netCDF4.py
index 661103b3..cd64c8be 100644
--- a/test/perf/benchmark-python-netCDF4.py
+++ b/test/perf/benchmark-python-netCDF4.py
@@ -5,6 +5,7 @@
 import netCDF4
 import numpy as np
 import timeit
+import sys
 
 def compute(v):
     tot = 0
@@ -13,9 +14,10 @@ def compute(v):
 
     return tot/v.shape[0]
 
-def process(fname):
-    with open("/proc/sys/vm/drop_caches","w") as f:
-        f.write("3")
+def process(fname,drop_caches):
+    if drop_caches:
+        with open("/proc/sys/vm/drop_caches","w") as f:
+            f.write("3")
 
     with netCDF4.Dataset(fname) as ds:
         v = ds["v1"]
@@ -24,15 +26,20 @@ def process(fname):
 
 
 if __name__ == "__main__":
+    drop_caches = "--drop-caches" in sys.argv
+
+    print("Python ",sys.version)
+    print("drop caches: ",drop_caches)
+
     fname = "filename_fv.nc";
-    tot = process(fname)
+    tot = process(fname,drop_caches)
 
     print("result ",tot)
 
     setup = "from __main__ import process"
     print("python-netCDF4 version ",netCDF4.__version__)
 
-    benchtime = timeit.repeat(lambda: process(fname), setup=setup,number = 1, repeat = 100)
+    benchtime = timeit.repeat(lambda: process(fname,drop_caches), setup=setup,number = 1, repeat = 100)
     with open("python-netCDF4.txt","w") as f:
         for bt in benchtime:
             print(bt,file=f)
diff --git a/test/perf/benchmark.sh b/test/perf/benchmark.sh
index e4a76c4b..e3acc2c3 100755
--- a/test/perf/benchmark.sh
+++ b/test/perf/benchmark.sh
@@ -1,7 +1,8 @@
 #!/bin/bash
 
-julia   benchmark-julia-NCDatasets.jl
-python3 benchmark-python-netCDF4.py
-Rscript benchmark-R-ncdf4.R
+args="$@"
+julia   benchmark-julia-NCDatasets.jl $args
+python3 benchmark-python-netCDF4.py $args
+Rscript benchmark-R-ncdf4.R $args
 
 julia summary.jl