Container performance notes

notes.rst

= Performance testing =

Kubernetes scaling SLO * 99% apiserver response under 1s * 99% pods scheduled under 5s (prepull images)

== Process performance ==

Containers vs VMs vx process on host

no CPU overhead KVM use few % when idle

http://domino.research.ibm.com/library/cyberdig.nsf/papers/0929052195DD819C85257D2300681E7B/$File/rc25482.pdf

== Storage performance ==

graphdriver Docker storage drivers have different penalties

Default driver (loop-lvm) is not suitable for production use, it is terrible http://www.projectatomic.io/blog/2015/06/notes-on-fedora-centos-and-docker-storage-drivers/

https://community.hortonworks.com/articles/87949/docker-storage-drivers-overview.html

production ready is: aufs, direct-lvm

https://docs.docker.com/engine/userguide/storagedriver/selectadriver/#future-proofing

== Scheduling performance ==

Kubemark https://github.com/kubernetes/community/blob/master/contributors/design-proposals/kubemark.md

CoreOS - Improving K8s scheduler performance https://coreos.com/blog/improving-kubernetes-scheduler-performance.html

=== Etcd tuning ===

Heartbeat interval max(avg(RTT)) Election timeout 5..10*Heartbeat

Periodicas snapshots: etcd append all key changes to log by default every 10k changes

etcd over SSL?

== Network performance ==

depends on driver/CNI default ~10us redirection latency