TCP (Client) Networking in Pure Bash

Recently I re-remembered about /dev/tcp - a virtual bash file system that allows you to directly connect to a remote TCP endpoint - without the use of nc / netcat / ncat.

While it only allows you to connect (no listening, sadly), it's still a great bash built-in that helps avoid awkward platform-specific issues.

Here's how you'd listen for a connection with netcat, sending as many random numbers as possible to the poor unsuspecting client:

netcat -l 0.0.0.0 6666 </dev/urandom

Here's how you'd traditionally connect to that via netcat:

netcat X.Y.Z.W 6666 | pv >/dev/null

The pv command there is not installed by default, but is a great tool that shows the amount of data flowing through a pipe. It's available in the repositories for most Linux distributions without any special configuration required - so sudo apt install pv should be enough on Debian-based distributions.

Now, let's look at how we'd do this with pure bash:

pv >/dev/null </dev/tcp/X.Y.Z.W/6666

Very neat! We've been able to eliminate an extra child process. The question is though: how do they compare performance-wise? Well, that depends on how we measure it. In my case, I measured a single connection, downloading data as fast as it can for 60 seconds.

Another test would be to open many connections and download lots of small files. While I haven't done that here, I theorise that the pure-bash method would win out, as it doesn't have to spawn lots of subprocesses.

In my test, I did this:

# Traditional method
timeout 60 nc X.Y.Z.W 6666 | pv >/dev/null
# Pure-Bash method
timeout 60 pv >/dev/null </dev/tcp/X.Y.Z.W/6666

The timeout command kills the process after a given number of seconds. The server I connected to was just this:

while true; do nc -l 0.0.0.0 6666 </dev/urandom; done

Running the above test, I got the following output:

$ timeout 60 pv >/dev/null </dev/tcp/172.16.230.58/6666
 652MiB 0:00:59 [11.2MiB/s] [                                      <=>         ]
$ timeout 60 nc 172.16.230.58 6666 | pv >/dev/null
 599MiB 0:01:00 [11.1MiB/s] [                                     <=>          ]

Method	Total Data Transferred
Traditional	599MiB
Pure Bash	652MiB

As it turns out, the pure bash method is apparently faster - by ~8.8%. I think this might have something to do with the lack of the additional sub-process, or some other optimisation that bash can apply when doing the TCP networking itself.

Found this interesting? Got a cool use for it? Discovered another awesome bash built-in? Comment below!