https://danpker.com/posts/faster-bash-startup/

Daniel Parker
Posts Tags RSS Github

Faster Bash Startup

2 Aug 2020

  * #bash

---------------------------------------------------------------------

My Bash start-up time is around 1.7 seconds. This doesn't sound like
it's too bad, but it feels like hours when you want to open a new
window to only run one command.

The first step is to run a profile of your Bash start-up, to try and
find out what exactly is taking so long.

How to profile a bash shell script slow startup?

This should give you a timestamp against each Bash command that gets
run, which you can use to work out which ones are taking the most
time.

For me, there are a few things that were highlighted.

    Note: Everything here will be specific to my setup (MacOS,
    Alacitty, Bash, Tmux) and so may and probably will differ.

I tested the difference between each change by using hyperfine.

hyperfine 'bash --login'

Dynamic paths

One of the big culprits for slowness is trying to be too clever. For
example, Homebrew has a prefix directory where it installs
everything, and environment variables that are set are set relative
to this directory.

For example, you might have some code that looks like this:

if [ -f $(brew --prefix)/etc/bash_completion ]; then
    . $(brew --prefix)/etc/bash_completion
fi

To allow for changing the default prefix directory, there is a
command brew --prefix that will resolve to where the prefix directory
is. However, this command can be quite slow and just replacing it
with a hardcoded path can give you a decent speed up:

Benchmark #1: bash prefix.sh
  Time (mean +- s):     289.9 ms +-   4.5 ms    [User: 178.9 ms, System: 103.6 ms]
  Range (min ... max):   285.4 ms ... 300.7 ms    10 runs

Benchmark #2: bash no_prefix.sh
  Time (mean +- s):     236.1 ms +-   2.4 ms    [User: 157.7 ms, System: 75.9 ms]
  Range (min ... max):   231.6 ms ... 241.8 ms    12 runs

Summary
  'bash no_prefix.sh' ran
    1.23 +- 0.02 times faster than 'bash prefix.sh'

This is even more significant if you have a prefix command that
specifies a particular library. E.g:

if [ -f $(brew --prefix openssl) ]; then
    OPENSSL_PREFIX="$(brew --prefix openssl)"
    export OPENSSL_INCLUDE_DIR="${OPENSSL_PREFIX}/include"
    export OPENSSL_LIB_DIR="${OPENSSL_PREFIX}/lib"
fi

In this case, using a hard coded path has a significant speed up.

Benchmark #1: bash prefix.sh
  Time (mean +- s):      1.095 s +-  0.015 s    [User: 627.0 ms, System: 444.6 ms]
  Range (min ... max):    1.079 s ...  1.131 s    10 runs

Benchmark #2: bash no_prefix.sh
  Time (mean +- s):       3.1 ms +-   0.4 ms    [User: 1.2 ms, System: 1.0 ms]
  Range (min ... max):     2.5 ms ...   5.4 ms    389 runs

  Warning: Command took less than 5 ms to complete. Results might be inaccurate.

Summary
  'bash no_prefix.sh' ran
  358.57 +- 47.07 times faster than 'bash prefix.sh'

Lesson: Avoid $(give_me_a_path) where possible by using a hard coded
path. Especially avoid brew --prefix <formula>.

Bash completions

The other major slow down in my Bash start-up was Bash completions.

Benchmark #1: source .profile
  Time (mean +- s):     208.6 ms +-   5.7 ms    [User: 90.0 ms, System: 115.6 ms]
  Range (min ... max):   203.7 ms ... 223.7 ms    13 runs

Benchmark #2: source .profile && source /usr/local/etc/bash_completion
  Time (mean +- s):     442.3 ms +-   3.3 ms    [User: 240.2 ms, System: 198.4 ms]
  Range (min ... max):   438.1 ms ... 449.2 ms    10 runs

Summary
  'source .profile' ran
    2.12 +- 0.06 times faster than 'source .profile && source /usr/local/etc/bash_completion'

As you can see above, adding completions doubles the start-up time of
my Bash session. I honestly don't know if I have ever even used them,
so it was an easy trade-off for me to make by removing them for a bit
more speed. If I had used them I would have thought more carefully
about it.

Conclusion

Using variations on the above I've managed to shave quite a bit of
time from my Bash startup (ignore the name change):

Benchmark #1: source ~/.bashrc
  Time (mean +- s):      1.669 s +-  0.025 s    [User: 924.4 ms, System: 706.6 ms]
  Range (min ... max):    1.644 s ...  1.728 s    10 runs

Benchmark #1: source ~/.profile
  Time (mean +- s):     216.2 ms +-   5.9 ms    [User: 94.2 ms, System: 119.5 ms]
  Range (min ... max):   210.2 ms ... 234.0 ms    13 runs

I could keep tweaking in an attempt to make it even faster, but
anything else at this point is likely to suffer from diminishing
returns. I mean, it has to take some time.

(c) 2010 - 2021 Daniel Parker