Measurements and comparisons of nabla containers
Clone or download

This is a repository containing the methodology for some experiments used to evaluate one aspect of container isolation: the attack surface to the host kernel. More specifically, these experiments measure how many kernel functions are accessed by an application as it runs and how complex those functions are.

The script runtest.bash performs a run of a test on one of the Docker containers that we have provided in the nabla-demo-apps repository. Currently this consists of:

Each run of runtest.bash sets up the kernel ftrace facility, runs the container, turns on tracing for the relevant pids, offers load, cleans up and processes the results. The raw and processed output ends up in a directory for later perusal.

The easiest way to replicate a test is to run commands of the form ./runtest.bash <RUNTIME> <CONTAINER> <OUTPUT_DIR>. RUNTIME can currently be runc (default docker), runnc (nabla), kata (kata containers), runsc (gvisor), or runsck (gvisor in kvm mode), and CONTAINER is one of those specified above. Here are some examples:

sudo ./runtest.bash runc nablact/node-express results/docker-node-express
sudo ./runtest.bash runnc nablact/node-express results/nabla-node-express
sudo ./runtest.bash kata nablact/node-express results/kata-node-express
sudo ./runtest.bash runsc nablact/node-express results/gvisor-node-express
sudo ./runtest.bash runsck nablact/node-express results/gvisork-node-express

We inform the Docker daemon of the alternate runtimes by adding them to /etc/docker/daemon.json (as also described in the runnc repository):

"runtimes": {
    "runnc": {
        "path": "/usr/local/bin/runnc"
    "kata": {
        "path": "/usr/bin/kata-runtime"
    "runsc": {
        "path": "/usr/local/bin/runsc"
    "runsck": {
        "path": "/usr/local/bin/runsc",
        "runtimeArgs": [


Technical notes

For ease of measurement, we do not measure the startup coverage of the containers, thus we need containers that are long-running enough to give the tracing enough time to turn on. Partially this is because of how we gather the relevant pids. This is done by looking through pstree for all the children of the docker-containe command specifying the -runtime-root, which is either runtime-runc (in the default case), runtime-runnc (in the nabla case), runtime-kata (in the kata containers case) or runtime-runsc (in the gvisor case). During the experiments, we pin all container runtime pids to core 0, to avoid confusing the function trace if a pid were to switch cores.

We also process the raw traces primarily to eliminate interrupts, which may perform work on behalf of other processes. There are three filtering programs, written in C, in the filters/ directory to aid in removing interrupt-related functions from the raw trace. See the README in that directory for more information.

Once filtering and processing is complete, a list of the unique functions called should appear in OUTPUT_DIR/trace.list. Counting the lines in this file will give a count of the unique functions. Alternately, grepping for functions starting with sys will show the system calls (e.g., grep -i "^sys\_")


To get a sense of how complex the unique functions are, we also perform a complexity analysis using the GNU complexity tool. We run complexity over the kernel version we used (sample output is included in this repository in the file named ./complexity). Then, we use those numbers to compute a total complexity sum for each case.

Our results

We have included some results for default docker, nabla, kata, and gvisor containers that were obtained using the graphs_generate.bash script.