github
/
k3s
mirror of https://github.com/k3s-io/k3s
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Update documentation about resource accounting

pull/6/head
Dr. Stefan Schimanski 2015-07-30 21:10:46 +02:00
parent a2fa41b73f
commit f3f617d9db
2 changed files with 21 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
contrib/mesos/docs/architecture.md
View File

@ -24,6 +24,20 @@ Kubernetes-Mesos uses "normal" Docker IPv4, host-private networking, rather than
![Network Diagram](networking.png)
## Resource Accounting
Mesos is designed to handle resource accounting and enforcement across the cluster. Part of that enforcement involves "growing" and "shrinking" the pool of resources allocated for executor containers.
The implementation of the k8sm-executor launches pods as Docker containers (just like the upstream kubelet). The containers are resource limited (cpu and memory) with the means of `docker run` by the kubelet code. Moreover, all containers launched by the kubelet code are children of the k8sm-executor cgroup. This parent cgroup is assigned to the k8sm-executor by the Mesos slave.
To actually enforce the defined resource limit for the k8sm-executor and its pods, enable the cpu and memory isolator in your Mesos slaves.
The described resource allocation also applies to static pods which are run on every Mesos slave which runs a k8sm-executor.
Kubernetes allows to define pods without resource limits for cpu and/or memory. The upstream kubelet will then run the containers without resource bounds. Because Mesos enforces resource accounting, it assign default container cpu and memory limits for those pods. By default these are 0.25 cpu shares and 64 MB of memory. These values can be customized via the `--default-container-cpu-limit` and `--default-container-mem-limit` of the k8sm-scheduler.
Note that currently static pods without cpu and memory limit are not allowed and will make the k8sm-scheduler refuse to start (compare the [k8sm issues](issues.md)).
[1]: http://mesos.apache.org/
[2]: https://issues.apache.org/jira/browse/MESOS-1806
[3]: issues.md#service-endpoints

16
contrib/mesos/docs/issues.md
View File

@ -7,15 +7,6 @@ Upon further consideration it has been decided that a greater alignment between
Currently it is not possible to specify pod placement constraints for the kubernetes-mesos scheduler.
This issue is being tracked here: https://github.com/mesosphere/kubernetes-mesos/issues/338
### Resource Allocation
Resource requirements (limits) specified on Kubernetes pods are currently ignored, both in the scheduler and on the node. Instead hardcoded values are used for the time being. This issue is being tracked here: https://github.com/mesosphere/kubernetes-mesos/issues/68.
In general Mesos is designed to handle resource accounting and enforcement across the cluster. Part of that enforcement involves "growing" and "shrinking" the pool of resources allocated for executor containers.
The current implementation of the kubelet-executor launches pods as Docker containers (just like the upstream kubelet) and makes no attempt to actually "contain" the pods that are launched. Because the kubernetes-mesos scheduler cannot depend on the kubelet-executor to properly contain resources, it foregoes implementing accurate resource accounting.
Recent changes to both the Docker and Kubernetes codebase have made it possible to implement the necessary changes in the kubelet-executor for proper pod containment. This is in the works and will be merged into a later version when ready.
### Ports
Mesos typically defines `ports` resources for each slave and these ports are consumed by tasks, as they are launched, that require one or more host ports.
@ -30,6 +21,13 @@ Ports declared outside that range (other than zero) will never match resource of
As opposed to Kubernetes proper, a missing pod container host port specification or a host port set to zero will allocate a host port from a resource offer.
### Static Pods
Static pods in general are supported by the k8sm-scheduler. The path of the pod definitions can be set via the `--static-pods-config` flag. There are two restrictions currently in that implementation:
- static pods *must have resource limits* on cpu and memory in their container specs (compare the [k8sm architecture](architecture.md))
- static pods *are read only once* by the k8sm-scheduler on startup. Only newly started executor will get the latest static pod specs from the defined static pod directory.
### Service Endpoints
At the time of this writing both Kubernetes and Mesos are using IPv4 addressing, albeit under different assumptions.