This is a PoC showing that a backend server with a L4 health check causes high CPU load when down compared to a backend server with no health check. We can reproduce that by starting two Vagrant VMs, install HaProxy, block the backend server IP via iptables and deploy a nearly identical config. The only difference is the check parameter on the server.

Install Vagrant and VirtualBox and run

vagrant up

Afterwards you have the high CPU load example on port 8080 and the low CPU load example on port 8081. You can find the status page on both ports on /stats.

Next log into the machines and start htop to monitor CPU load:

vagrant ssh haproxy-high-cpu
sudo htop

vagrant ssh haproxy-low-cpu
sudo htop

Next you can start to put some load on the example instances. In this case we use the tool wrk but any load testing tool should be fine.

watch wrk http://<host-ip>:8080 -c 100 -t 4 -d 120

watch wrk http://<host-ip>:8081 -c 100 -t 4 -d 120

It is recommended to do this from another machine since wrk uses a lot of CPU itself.

Now you should be able to see that the high CPU load example indeed uses a lot more CPU than the other one. The only difference is the check parameter on the backend's server. You can find the config files at the root of this repo.