Eliminating the impact of UnionFS on results when benchmarking inside Docker

Question

I am trying to benchmark the overall system performance of running Docker using the Phoronix Test Suite 6.4.0 Milestone 2 running inside a fedora:23 image based container.

One thing that must be considered is, that Docker uses a proprietary UnionFS to store data. However, when running a real-world application (like Apache) inside Docker, the persistent data are usually stored on a dedicated folder on the host, running on a standard Linux filesystem like ext4, or in my case btrfs.

The solution I propose is to use the "docker volume" to mount a host directory into docker. The thing is I don't know which directories are to be used in the benchmarks and have to be mounted inside the Docker container.

The test Suite pts/disk for example should definitely use docker volumes instead of UnionFS. It contains these tests.

pts/compress-gzip
pts/sqlite
pts/apache
pts/pgbench
pts/compilebench
pts/iozone
pts/dbench
pts/fs-mark
pts/fio
pts/tiobench
pts/postmark
pts/aio-stress
pts/unpack-linux

Which directories in the Docker container should be mounted from host (made docker volumes)? Is it even a good idea to use the docker volumes? Are there any other caveats to consider when benchmarking Docker?

Answer 1

As mentioned in "Mount a shared-storage volume as a data volume"

In addition to mounting a host directory in your container, some Docker volume plugins allow you to provision and mount shared storage, such as iSCSI, NFS, or FC.

Since docker 1.9, this is done with docker volume create (and then you mount those volumes in their expected paths in the container)

That allows you to:

• make those volumes independent of the host
• still bypass UFS
• experiments with different Docker volume plugins
• compare their performance benchmarks

Which directories in the Docker container should be mounted from host

Any folder which include persistent data, and/or has an high volume of IO operations.