Benchmarking the overhead of calling a function directly a vs using a wrapping function like #(a).
In Clojure, you would do this in order to make the call more "REPL friendly", as in the latter case redefining a will change the behavior without requiring a restart.
I’ve had several conversations about this where performance was raised as a concern, so I decided to measure the overhead.
To me, the difference appears insignificant.
| No compiler options | Direct Linking | |
|---|---|---|
Indirect use |
Evaluation count : 4845042720 in 60 samples of 80750712 calls. Execution time mean : 9.534677 ns Execution time std-deviation : 0.372420 ns Execution time lower quantile : 9.066588 ns ( 2.5%) Execution time upper quantile : 10.304019 ns (97.5%) Overhead used : 3.159809 ns |
Evaluation count : 5206759680 in 60 samples of 86779328 calls. Execution time mean : 4.784659 ns Execution time std-deviation : 0.332266 ns Execution time lower quantile : 4.282720 ns ( 2.5%) Execution time upper quantile : 5.414823 ns (97.5%) Overhead used : 7.177589 ns |
Direct use |
Evaluation count : 4421273100 in 60 samples of 73687885 calls. Execution time mean : 10.696803 ns Execution time std-deviation : 0.308295 ns Execution time lower quantile : 10.324632 ns ( 2.5%) Execution time upper quantile : 11.267389 ns (97.5%) Overhead used : 3.159809 ns |
Evaluation count : 4568565000 in 60 samples of 76142750 calls. Execution time mean : 6.339981 ns Execution time std-deviation : 0.395668 ns Execution time lower quantile : 5.724480 ns ( 2.5%) Execution time upper quantile : 7.023956 ns (97.5%) Overhead used : 7.177589 ns |
You can run the tests yourself with ./run.
Some things to note:
-
You will need the Clojure CLI tools installed
-
I disable function inlining in the JVM options in an attempt to prevent certain optimizations happening because this is a toy example and I don’t want the simpleness of the function to be ruined.
./rundoes show output with both inlining disabled & enabled though. -
It will run 2 JVMs, one after the other.