Slides for the docker meetup talk on overlay networks in Docker.

1. Network Overlay Options in
Docker
Syed Mushtaq Ahmed
syed.ahmed4@mail.mcgill.ca

2. Networking in Docker
● When docker daemon starts, it sets up the docker0 bridge which is the entry point
to all container traffic.
● Communication between local containers works but anything outside should be
port forwarded.
● Can cause problems if multiple containers want to communicate over the same
port.
● Overlay networks allow the possibility of seamless communication between
multiple containers without jumping multiple hoops.
● We examine three overly networking options that are available in Docker. Weave,
Flannel, Libnetwork.

3. Setup

4. Weave
● “Weave creates a virtual network that connects Docker containers deployed
across multiple hosts and enables their automatic discovery.”[1]
● Weave creates a custom bridge to which each container connects.
● Uses a “router” container which intercepts packets destined to the bridge,
encapsulates them and sends it over to the right peer router.
● Each router learns which mac addresses belong to which peer router and is also
aware of the overall topology.
● Uses a custom encapsulation format and batches multiple frames in a single UDP
payload
[1] https://github.com/weaveworks/weave#readme

5. Weave Setup
#install
curl -L git.io/weave -o /usr/local/bin/weave
chmod a+x /usr/local/bin/weave
#start
weave launch [$PEER_IP]
eval $(weave env)
#run
docker run --name c1 -it ubuntu

6. Weave
Overlay Throughput
Native (Inter-VM) 3.2 Gb/s
Weave 147 Mb/s
Weave is slow because the router container uses PCAP to capture packets and encapsulate
them in userspace which is every expensive.

7. Flannel
● “Flannel is a virtual network that gives a subnet to each host for use with
container runtimes.”[2]
● Each host gets a subnet assigned to it and IP allocations to a container happen
from that subnet.
● Uses etcd for storing configuration.
● Can have multiple “backends” (UDP, VxLAN, AWS-VPC)
● docker0 is kept the default bridge so no extra interfaces in the container.
● Supports multi-network mode but is static and still experimental.
[2] https://github.com/coreos/flannel

8. # Setup Etcd ...
#Build flannel
git clone https://github.com/coreos/flannel.git
cd flannel
docker run -v `pwd`:/opt/flannel -i -t google/golang /bin/bash -c "cd /opt/flannel &&
./build"
#push network config to etcd
curl -L http://127.0.0.1:2379/v2/keys/coreos.com/network/config -XPUT -d value='{
"Network": "10.0.0.0/8",
"SubnetLen": 20,
"SubnetMin": "10.10.0.0",
"SubnetMax": "10.99.0.0",
"Backend": {
"Type": "vxlan",
"Port": 7890
}
}'
Flannel Setup (kernel > 3.15)

9. #start flannel
cd flannel/bin
./flanneld -etcd-endpoints="http://127.0.0.1:2379"
# start docker with the flannel (you may have to change the docker0's IP
service docker stop
source /run/flannel/subnet.env
docker -d --bip=${FLANNEL_SUBNET} --mtu=${FLANNEL_MTU}
#start containers normally
docker run -it ubuntu
Flannel Setup (kernel > 3.15)

10. Flannel
Overlay Throughput
Native (Inter-VM) 3.2 Gb/s
Weave 147 Mb/s
Flannel 1.22 Gb/s
Flannel is faster than Weave because it uses the kernel Vxlan driver thus avoiding packet copy to
user space.

11. Libnetwork
● Currently in active development.
● Integrated tightly with Docker, it provides native multi-host networking.
● Flexible to support any external drivers (Weave for example).
● Defines networks and services as top level objects.
● We can dynamically create multiple networks, services belonging to different
networks and attach them to containers.

12. #build docker binary with experimental support
git clone https://github.com/docker/docker.git; cd docker
DOCKER_EXPERIMENTAL=1 make
#setup a Key-Value store (using Consul here)
#host1
consul agent -server -bootstrap -data-dir /tmp/consul -bind=<host-1-ip-address>
#host2
consul agent -data-dir /tmp/consul -bind <host-2-ip-address>
consul join <host-1-ip-address>
#start docker
docker -d --kv-store=consul:localhost:8500 --
label=com.docker.network.driver.overlay.bind_interface=eth0
[--label=com.docker.network.driver.overlay.neighbor_ip=<host-1-ip-address>]
Libnetwork Setup (kernel > 3.15)

13. #Create network with overlay driver
docker network create -d overlay mynet
#create a service under the network
#host1
docker service publish svc1.mynet
#host2
docker service publish svc2.mynet
#start a container and attach the service to it
#host1
CID=$(docker run -itd ubuntu)
docker service attach $CID svc1.mynet
#host2
CID=$(docker run -itd ubuntu)
docker service attach $CID svc2.mynet
Libnetwork Setup (kernel > 3.15)

14. Libnetwork
Overlay Throughput
Native (Inter-VM) 3.2 Gb/s
Weave 147 Mb/s
Flannel 1.22 Gb/s
Libnetwork 1.32 Gb/s
Libnetwork uses the same Vxlan driver as Flannel. It has a slightly higher throughput possibly
because Flannel sets a slightly lower MTU (1450 instead of 1500) on the docker bridge.

15. Other approaches
● Rancher uses IPSec tunnels between different hosts to implement their overlay.
● Socketplane used Open VSwitch as their container bridge and used its VxLAN
tunneling capability.

16. Questions?