Improving gVisor Memory Subsystem Performance

This is a course project. See our final report. Read on if you would like to reproduce our results.

Environment Setup

Throughout the project, the target OS is assumed to be Ubuntu 18.04 LTS (Bionic Beaver). To prepare a fresh OS for following work, just do:

./script/provision-ubuntu.sh

Then you could check the versions of major tools in use:

./script/check-versions.sh

Vagrant

gVisor only works on Linux. If your local OS is not Linux, you could use a virtual machine for local development. A Vagrantfile is provided to facilitate setting up the VM environment. Just do:

$ vagrant up

This project directory is synced with ~/work within the VM. To get to work:

$ vagrant ssh
$ cd work

GCP and AWS

The cloud environment is not so much different from a local one. Start a proper instance, clone this repository somewhere, run the provision script, and you shall be good to go.

Benchmark Run

Once the environment is set up, we could move on to do benchmarks. Following operations are assumed to be invoked in the benchmark directory.

Manual commands

First compile the source code:

$ make

The compiled binaries could be found in bin. To run a benchmark natively:

$ ./bin/malloc_benchmark 1000 1024 0

To move forward, we need to build a docker image:

$ docker image build . --tag gvisor-mem-perf

To run the same command in Docker:

$ docker run gvisor-mem-perf ./bin/malloc_benchmark 1000 1024 0

To run the same command in Docker with gVisor:

$ docker run --runtime=runsc gvisor-mem-perf ./bin/malloc_benchmark 1000 1024 0

Other runtimes are configured in daemon.json. The usages of *-debug and *-prof runtimes are more advanced. The details could be found in gVisor documentation on debugging.

Automation scripts

To run a whole suite of benchmarks:

$ ./run_all.sh

Data Analysis

In the previous step, the results are saved in benchmark/data. To persist the data for further analysis, we need to move that directory to be a subdirectory of data. The naming convention is data/{machine_type}. For example, data/gcp-n1-standard-4 means results from a n1-standard-4 instance on GCP.

Once the data directory is populated, we could visualize the results using a Python script:

$ ./data/plot.py

qobilidop / gvisor-mem-perf