Testing Docker multi-host network performanceVadim Tkachenko
For this set of tests, I wanted to focus more on Docker networking using multiple hosts. Mostly because when we set up a high availability (HA) environment (using Percona XtraDB Cluster, for example) the expectation is that instances are running on different hosts.
Another reason for this test is that Docker recently announced the 1.12 release, which supports Swarm Mode. Swarm Mode is quite interesting by itself — with this release, Docker targets going deeper on Orchestration deployments in order to compete with Kubernetes and Apache Mesos. I would say Swarm Mode is still rough around the edges (expected for a first release), but I am sure Docker will polish this feature in the next few releases.
Swarm Mode also expects that you run services on different physical hosts, and services are communicated over Docker network. I wanted to see how much of a performance hit we get when we run over Docker network on multiple hosts.
Network performance is especially important for clustering setups like Percona XtraDB Cluster and MySQL Group Replication (which just put out another Lab release).
For my setup, I used two physical servers connected over a 10GB network. Both servers use 56 cores total of Intel CPUs.
Sysbench setup: data fits into memory, and I will only use primary key lookups. Testing over the network gives the worst case scenario for network round trips, but it also gives a good visibility on performance impacts.
The following are options for Docker network:
- No Docker containers (marked as “direct” in the following results)
- Docker container uses “host” network (marked as “host”)
- Docker container uses “bridge” network, where service port exposed via port forwarding (marked as “bridge”)
- Docker container uses “overlay” network, both client and server are started in containers connected via overlay network (marked as “overlay” in the results). For “overlay” network it is possible to use third-party plugins, with different implementation of the network, the most known are:
- Calico network https://github.com/projectcalico/calico-containers
- Weave network https://github.com/weaveworks/weave
For multi-host networking setup, only “overlay” (and plugins implementations) are feasible. I used “direct”, “host” and “bridge” only for the reference and as a comparison to measure the overhead of overlay implementations.
The results I observed are:
|Client||Server||Throughput, tps||Ratio to “direct-direct”|
|Calico overlay||Calico overlay||246202||0.87|
|Weave overlay||Weave overlay||11554||0.044|
- “Bridge” network added overhead, about 12%, which is in line with my previous benchmark. I wonder, however, if this is Docker overhead or just the Linux implementation of bridge networks. Docker should be using the setup that I described in Running Percona XtraDB Cluster nodes with Linux Network namespaces on the same host, and I suspect that the Linux network namespaces and bridges add overhead. I need to do more testing to verify.
- Native “Overlay” Docker network struggled from performance problems. I observed issues with ksoftirq using 100% of one CPU core, and I see similar reports. It seems that network interruptions in Docker “overlay” are not distributed properly across multiple CPUs. This is not the case with the “direct” and “bridge” configuration. I believe this is a problem with the Docker “overlay” network (hopefully, it will eventually be fixed).
- Weave network showed absolutely terrible results. I see a lot of CPU allocated to “weave” containers, so I think there are serious scalability issues in their implementation.
- Calico plugin showed the best result for multi-host containers, even better than “bridge-bridge” network setup
If you need to use Docker “overlay” network — which is a requirement if you are looking to deploy a multi-host environment or use Docker Swarm mode — I recommend you consider using the Calico network plugin for Docker. Native Docker “overlay” network can be used for prototype or quick testing cases, but at this moment it shows performance problems on high-end hardware.