Skip to content

Commit

Permalink
Updated CPU manager docs to match implementation. (#5332)
Browse files Browse the repository at this point in the history
* Noted limitation of alpha static cpumanager.

* Updated CPU manager docs to match implementation.

- Removed references to CPU pressure node condition and evictions.
- Added note about new --cpu-manager-reconcile-period flag.
- Added note about node allocatable requirements for static policy.
- Noted limitation of alpha static cpumanager.

* Move cpu-manager task link to rsc mgmt section.
  • Loading branch information
ConnorDoyle authored and steveperry-53 committed Sep 12, 2017
1 parent 2cf12d2 commit a1dfece
Show file tree
Hide file tree
Showing 2 changed files with 34 additions and 26 deletions.
2 changes: 1 addition & 1 deletion _data/tasks.yml
Original file line number Diff line number Diff line change
Expand Up @@ -124,6 +124,7 @@ toc:
- docs/tasks/administer-cluster/quota-pod-namespace.md
- docs/tasks/administer-cluster/quota-api-object.md
- docs/tasks/administer-cluster/opaque-integer-resource-node.md
- docs/tasks/administer-cluster/cpu-management-policies.md
- docs/tasks/administer-cluster/access-cluster-api.md
- docs/tasks/administer-cluster/access-cluster-services.md
- docs/tasks/administer-cluster/securing-a-cluster.md
Expand All @@ -140,7 +141,6 @@ toc:
- docs/tasks/administer-cluster/cpu-memory-limit.md
- docs/tasks/administer-cluster/out-of-resource.md
- docs/tasks/administer-cluster/reserve-compute-resources.md
- docs/tasks/administer-cluster/cpu-management-policies.md
- docs/tasks/administer-cluster/guaranteed-scheduling-critical-addon-pods.md
- docs/tasks/administer-cluster/declare-network-policy.md
- title: Install Network Policy Provider
Expand Down
58 changes: 33 additions & 25 deletions docs/tasks/administer-cluster/cpu-management-policies.md
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,7 @@ directives.
## CPU Management Policies

By default, the kubelet uses [CFS quota](https://en.wikipedia.org/wiki/Completely_Fair_Scheduler)
to enforce pod CPU limits.  When the node runs many CPU bound pods,
to enforce pod CPU limits.  When the node runs many CPU-bound pods,
the workload can move to different CPU cores depending on
whether the pod is throttled and which CPU cores are available at
scheduling time.  Many workloads are not sensitive to this migration and thus
Expand All @@ -25,13 +25,25 @@ However, in workloads where CPU cache affinity and scheduling latency
significantly affect workload performance, the kubelet allows alternative CPU
management policies to determine some placement preferences on the node.

Enable these management policies with the `--cpu-manager-policy` kubelet
option.  There are two supported policies:
### Configuration

* `none`: the default, which represents the existing scheduling behavior
The CPU Manager is introduced as an alpha feature in Kubernetes v1.8. It
must be explicitly enabled in the kubelet feature gates:
`--feature-gates=CPUManager=true`.

The CPU Manager policy is set with the `--cpu-manager-policy` kubelet
option. There are two supported policies:

* `none`: the default, which represents the existing scheduling behavior.
* `static`: allows pods with certain resource characteristics to be
granted increased CPU affinity and exclusivity on the node.

The CPU manager periodically writes resource updates through the CRI in
order to reconcile in-memory CPU assignments with cgroupfs. The reconcile
frequency is set through a new Kubelet configuration value
`--cpu-manager-reconcile-period`. If not specified, it defaults to the same
duration as `--node-status-update-frequency`.

### None policy

The `none` policy explicitly enables the existing default CPU
Expand All @@ -49,8 +61,13 @@ using the [cpuset cgroup controller](https://www.kernel.org/doc/Documentation/cg
**Note:** System services such as the container runtime and the kubelet itself can continue to run on these exclusive CPUs.  The exclusivity only extends to other pods.
{: .note}

**Note:** The alpha version of this policy does not guarantee static
exclusive allocations across Kubelet restarts.
{: .note}

This policy manages a shared pool of CPUs that initially contains all CPUs in the
node minus any reservations by the kubelet `--kube-reserved` or
node. The amount of exclusively allocatable CPUs is equal to the total
number of CPUs in the node minus any CPU reservations by the kubelet `--kube-reserved` or
`--system-reserved` options. CPUs reserved by these options are taken, in
integer quantity, from the initial shared pool in ascending order by physical
core ID.  This shared pool is the set of CPUs on which any containers in
Expand All @@ -59,26 +76,21 @@ cpu `requests` also run on CPUs in the shared pool. Only containers that are
both part of a `Guaranteed` pod and have integer CPU `requests` are assigned
exclusive CPUs.

**Note:** When reserving CPU with `--kube-reserved` or `--system-reserved` options, it is advised to use *integer* CPU quantities.
**Note:** The kubelet requires a CPU reservation greater than zero be made
using either `--kube-reserved` and/or `--system-reserved` when the static
policy is enabled. This is because zero CPU reservation would allow the shared
pool to become empty.
{: .note}

As `Guaranteed` pods whose containers fit the requirements for being statically
assigned are scheduled to the node, CPUs are removed from the shared pool and
placed in the cpuset for the container.  CFS quota is not used to bound
placed in the cpuset for the container. CFS quota is not used to bound
the CPU usage of these containers as their usage is bound by the scheduling domain
itself. In others words, the number of CPUs in the container cpuset is equal to the integer
CPU `limit` specified in the pod spec.  This static assignment increases CPU
affinity and decreases context switches due to throttling for the CPU bound
CPU `limit` specified in the pod spec. This static assignment increases CPU
affinity and decreases context switches due to throttling for the CPU-bound
workload.

In the event that the shared pool is depleted the kubelet takes two actions:

* Evict all pods that include a container that does not specify a `cpu`
quantity in `requests` as those pods now have no CPUs on which to run.
* Set a `NodeCPUPressure` node condition to `true` in the node status. When
this condition is true, the scheduler will not assign any pod to the node
that has a container which lacks a `cpu` quantity in `requests`.

Consider the containers in the following pod specs:

```yaml
Expand All @@ -89,8 +101,7 @@ spec:
```
This pod runs in the `BestEffort` QoS class because no resource `requests` or
`limits` are specified. It is evicted if shared pool is depleted. It runs
in the shared pool.
`limits` are specified. It runs in the shared pool.

```yaml
spec:
Expand All @@ -105,9 +116,8 @@ spec:
```

This pod runs in the `Burstable` QoS class because resource `requests` do not
equal `limits` and the `cpu` quantity is not specified. It is
evicted if shared pool is depleted. It runs in the shared pool.

equal `limits` and the `cpu` quantity is not specified. It runs in the shared
pool.

```yaml
spec:
Expand All @@ -124,9 +134,7 @@ spec:
```

This pod runs in the `Burstable` QoS class because resource `requests` do not
equal `limits`. The non-zero `cpu` quantity in `requests` prevents the
shared pool from depleting. It runs in the shared pool.

equal `limits`. It runs in the shared pool.

```yaml
spec:
Expand Down

0 comments on commit a1dfece

Please sign in to comment.