Benchmark Tools for BPF Tail Calls

This repository contains an Ansible playbook and the sources of a Docker image to execute BPF tail call benchmarks on remote servers. It measures the cost of tail call chains of varying lengths using kern.c as the BPF program. The Docker image is available at docker.io/pchaigno/tail-calls:v1.0.

Results produced by these tools are presented at https://pchaigno.github.io/ebpf/2021/03/22/cost-bpf-tail-calls.html.

Requirements

The Ansible playbook was developed to work with Packet's Ubuntu 20.04 and may not work with other distributions. Only Ansible and the kernel debian packages are required on the machine used to run the playbook.

Running the Benchmark

Populate hosts.ini

You need to complete hosts.ini with the IP addresses of your benchmark servers, a string matching the kernel debian package to install, and the file where results should be written. For example:

147.75.100.157 kernel=5.10.0-retpoline boot_order=0 output=results-5.10-retpoline.txt ansible_user=root ansible_python_interpreter=/usr/bin/python3
147.75.100.157 kernel=5.4.0-retpoline boot_order=2 output=results-5.4-retpoline.txt ansible_user=root ansible_python_interpreter=/usr/bin/python3

The kernel string will be used in the file glob linux-*[kernel]*.deb to select the debian packages to install. boot_order should be set to 0 on kernels 5.5+ and 2 on others.

Execute the Ansible Playbook

You can then execute ansible-playbook:

ansible-playbook -v -i hosts.ini -e nb_runs=100 -e nb_measures=10 -e result_dir=/home/paul/pchaigno.github.io playbook.yaml

The Ansible tasks are executed on all servers in parallel without waiting for other servers.

• result_dir is the directory where the output result files mentioned above will be saved.
• nb_runs corresponds to the number of time each BPF program will be executed in the kernel. Only the mean of those nb_runs executions will be reported.
• nb_measures corresponds to the number of times BPF_PROG_TEST_RUN will be called. With the above parameters, each tail call chain will run a thousand times and 10 means will be reported.

License

This project is under MIT license.