Kube-proxy requests 2x cpu shares of addons. #22022
Conversation
|
@kubernetes/goog-cluster |
|
Labelling this PR as size/XS |
k8s-github-robot
added
the
size/XS
| @@ -31,6 +31,9 @@ spec: | |||
| containers: | |||
| - name: kube-proxy | |||
| image: {{pillar['kube_docker_registry']}}/kube-proxy:{{pillar['kube-proxy_docker_tag']}} | |||
| resources: | |||
| requests: | |||
There was a problem hiding this comment.
We should set limits ideally.
There was a problem hiding this comment.
I don't want to curtail kube-proxy, I want it to get what it needs (plus I'm not 100% sure how much it will need).
But I guess it doesn't make sense giving unlimited cpu to the proxy if nothing can run on the node, and if it bumps into the limit it's because something else with higher priority came along, and it'll just be slow, unlike memory.
There was a problem hiding this comment.
requests are OK for now I guess.
Depending on where we go with the QoS policies, this might make kube-proxy
have less priority. But given that it is a DaemonSet, I guess the policies
will be thwarted anyways.
On Thu, Feb 25, 2016 at 3:25 PM, Prashanth B notifications@github.com
wrote:
In cluster/saltbase/salt/kube-proxy/kube-proxy.manifest
#22022 (comment)
:@@ -31,6 +31,9 @@ spec:
containers:
- name: kube-proxy
image: {{pillar['kube_docker_registry']}}/kube-proxy:{{pillar['kube-proxy_docker_tag']}}
- resources:
requests:I don't want to curtail kube-proxy, I want it to get what it needs (plus
I'm not 100% sure how much it will need).But I guess it doesn't make sense giving unlimited cpu to the proxy if
nothing can run on the node, and if it bumps into the limit it's because
something else with higher priority came along, and it'll just be slow,
unlike memory.—
Reply to this email directly or view it on GitHub
https://github.com/kubernetes/kubernetes/pull/22022/files#r54182925.
There was a problem hiding this comment.
Yeah none of the master components set limits, for the same reason. If kube-proxy on a node is really slow it could impact a cluster spanning service, so I think we should leave it as request for now and revisit if required.
bprashanth
force-pushed
the
proxy_cpu_request
branch
2 times, most recently
from
98acea3 to
1834834
Compare
bprashanth
force-pushed
the
proxy_cpu_request
branch
from
1834834 to
7d47d2d
Compare
vishh
added
the
lgtm
|
GCE e2e build/test passed for commit e34bbafd8962dcda8d7f12845e11607eccea4570. |
bprashanth
added
the
priority/important-soon
|
GCE e2e build/test failed for commit 7d47d2d. Please reference the list of currently known flakes when examining this failure. If you request a re-test, you must reference the issue describing the flake. |
|
GCE e2e build/test passed for commit 98acea32d0fd69242024073edb8f1ff269d3daff. |
|
GCE e2e build/test failed for commit 1834834269333cdd07bd2b0d80366e6ba116a473. Please reference the list of currently known flakes when examining this failure. If you request a re-test, you must reference the issue describing the flake. |
|
GCE e2e build/test passed for commit 7d47d2d. |
|
@k8s-oncall this should help with services test flake, if the merge bot doesn't get to it overnight |
|
Merging manually - this doesn't seem risky. |
Kube-proxy requests 2x cpu shares of addons.
|
Actually - it appeared that it broke kubemark, because we haven't had enough place to run hollow-nodes in the cluster. But this is already fixed. @gmarek |
|
I am shocked that kube-proxy needs more CPU. It barely does anything. Do
|
|
I am sure if @bprashanth actually profiled kube-proxy before making this PR ? |
|
I am sure this was not a wild guess, I am just surprised to believe it needs that much CPU |
|
in other words - we need to figure out why. RAM I can buy. CPU, no way |
|
Prashant can elaborate on how he reached the more-cpu-needed conclusion Sent from my iPhone
|
|
It doesn't need 200m, we need to figure out how much. It surely was starved, it was getting as many shares as pause, see: #21820 (comment). 200m is just a number > than the addons. |
|
Sure, we can fine-tune it later On Mon, Feb 29, 2016 at 10:00 AM, Prashanth B notifications@github.com
|
see #21820 (comment)
@kubernetes/goog-node 200m should be fine right?