MQTT protocol bench test based on Docker simulated sensor and Eclipse Mosquitto Broker

Objective of the bench is to estimate the typical bandwitdh need of a MQTT protocol implementation. Th structure is based on a virtual net composed of sensor and sinks (as publishers), a broker based on Mosquitto and a series of client (as subscribers).

Based on minimal Alpine image that run a C program simulating the sensor behaviour. Also a python and Java implementation, but C was choosen for resource reason.

• Python: Apline apk installation require almost a GigaByte on disc
• Java: 20 MByte container RAM usage due to virtual machine
• C implementation: 0,9 MByte (!!!) container RAM occupation. 100 clients takes only 100MByte

The C implementation use the MQTT-C library: https://github.com/LiamBindle/MQTT-C/

Alpine with mosquitto running on a Virtualbox virtual machine due to IP network reason. With Wireshark is easier to capture and classify traffic.

Sinks is some simulated node that aggregate MQTT message in the same TCP payload. This makes the MQTT be less resources-angry due to dramatic decrease in protocol header overhead.

Before everything, change the IP to which client connects in file clients/C/publisher/simple_publisher.c Also change other settings, like QoS level, etc...

Compile

cd build  
cmake ..  
make  
mv publish ../../  

The last command move the new executable in the folder of Dockerfile, which will copy it on the container
The executable takes three parameter in:

./publish {boolean} {boolean} {integer} {integer}

args 1: 0 for a message every second, 1 for realistic random behaviour
args 2: 0 if is a client, 1 for sink functionalities
agrs 3: number of client behind a sink
agrs 4: payload length to send random string

To start the docker's client:

docker-compose up -d