Hi Larry,
Great work. I really like the approach. I'll try it out asap.
Garth
On Thu, 12 Jun, 2014 at 6:39 AM, Lizao Li <[email protected]> wrote:
I made a dockerized version of FEniCS (to be defined below). I would
like to hear some comments and see if there is interest in adding
this as yet another way of deploying FEniCS.
Docker (http://www.docker.com/) is an open source lightweight virtual
machine, popular among the cloud computing community. It has many
advanced features but the only thing that matters here is that it is
easy, fast, and runs almost everywhere (win, mac, linux, google
cloud, amazon ec2, etc, but no, not yet android~).
The result is a "FEniCS-in-a-box" with the following objectives
achieved:
- Get FEniCS running with <5 lines of command on most computers (I
get FEniCS with 3 lines on ArchLinux).
- Preferrably not compiling for hours. If it has to be compiled, it
is guaranteed to succeed (even on a cluster with weird toolchains).
- FEniCS is properly contained in a stable environment (unaffected by
things like boost, gcc updates) and can be shared between computers
(for checking and comparison).
- Minimal loss of performance.
To test it:
1. Go to http://docs.docker.com, find out how to install docker on
your system (1 line command in ArchLinux, similar for most linux
distros).
2. Deploy the FEniCS 1.4 image (1 line, download ~4GB, pretty fast,
no compiling)
sudo docker pull lzlarryli/fenics:test
2'. Alternatively, you can build FEniCS against a minimum Ubuntu (2
lines, ~1Gb download, ~30 min compiling):
wget http://www.math.umn.edu/~lixx1445/fenics_docker/Dockerfile
sudo docker build -t="lzlarryli/fenics:test" .
3. Run FEniCS (1 line):
sudo docker run -t -i --name fenics -v /tmp:/tmp
lzlarryli/fenics:test /bin/bash
The above creates a virtual machine instance based on the FEniCS
image:
-t means attach a terminal
-i means interactive
--name gives a name to the machine
-v /tmp:/tmp mounts the local /tmp to the /tmp in the
virtual machine so files can be shared.
After running that line, you end up in the FEniCS demo directory and
can go ahead and run some demos (fenics.conf is souced already).
4. Once done, run exit. It is the same as ssh into another computer.
To use FEniCS again, just run
sudo docker start fenics
sudo docker attach fenics
[ctrl+c]
Clean up
Remove this virtual machine instance:
sudo docker rm fenics
Remove the downloaded image:
sudo docker rmi lzlarryli/fenics:test
Uninstall docker.
How is this build?
It is built from a simple Dockerfile
(http://www.math.umn.edu/~lixx1445/fenics_docker/Dockerfile), which
basically downloads a minimal Ubuntu (~80Mb) and then use Dorsal to
compile a FEniCS in it. In fact with the same file, DockerHub (also
free) can link repositories to Bitbucket, which completely automates
the process of building images. DockerHub is free for binary
repositoreis of arbitrary size, as long as it is public. This is one
of Docker's advantage over other virtual machines, where images have
to be produced by hand and hosted somewhere.
Sudo?
It is annoying that sudo is used all the time. This undesirable
effect can be dealt with by installing sshd in the virtual machine
and expose the port to the local machine (2 lines, see Docker docs).
Then administrators (who not necessarily know how to compile FEniCS)
can use sudo to install docker, FEniCS, and run the virtual machine
as a daemon. Users can then ssh into it without sudo or anything
else. The users are contained in the virtual machine, so it is safe.
Performance?
This is the part I was not able to test in a reasonable way. I have
FEniCS 1.3 on various systems (none of them runs ubuntu) and FEniCS
1.4 on the virtual Ubuntu machine. The dockerized version runs
enlarged FEniCS demo problems faster than the native on the same
computer, which is ridiculous. I can only say that probably 1.4 is
faster than 1.3 (good job, btw). Most modern (reads
non-graduate-student-office) computers have virtualization, which
makes the virtual machine performance comparable to the native one.
You are welcome to test and show benchmark results.
Trade-offs?
Precompiled binaries or binaries compiled against standard libraries
certainly performs worse than those compiled against
hardware-specific optimized libraries. Yet there are quite many cases
where performance is not critical but easy deployment is: (1)
students, (2) new users, (3) code developement and testing, (4)
cluster/system admins who are not familar with numerical software,
(5) theoretical numerical analysts (use cluster only when memory
outflows), (6) people who have to run FEniCS on many different
platforms. Personally, I am (5)(6) and sometimes (4) breaks FEniCS on
machines I want to use. This at least makes my life quite easier.
More info about docker (can be interesting):
http://www.docker.com/tryit/
https://docs.docker.com/userguide/
--
Lizao (Larry) Li
Univeristy of Minnesota
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics
