Bug 1814048

Summary: Default openshift install requests too many CPU resources to install all components, requests of components on cluster are wrong
Product: OpenShift Container Platform Reporter: Alay Patel <alpatel>
Component: InstallerAssignee: Stephen Benjamin <stbenjam>
Installer sub component: openshift-installer QA Contact: Amit Ugol <augol>
Status: CLOSED DUPLICATE Docs Contact:
Severity: urgent    
Priority: urgent CC: adahiya, ccoleman, esimard, rphillips, sdodson, wjiang, zyu
Version: 4.4   
Target Milestone: ---   
Target Release: 4.4.0   
Hardware: Unspecified   
OS: Unspecified   
Whiteboard:
Fixed In Version: Doc Type: If docs needed, set a value
Doc Text:
Story Points: ---
Clone Of: 1812583 Environment:
Last Closed: 2020-03-17 13:20:06 UTC Type: ---
Regression: --- Mount Type: ---
Documentation: --- CRM:
Verified Versions: Category: ---
oVirt Team: --- RHEL 7.3 requirements from Atomic Host:
Cloudforms Team: --- Target Upstream Version:
Embargoed:
Bug Depends On: 1812583    
Bug Blocks:    

Description Alay Patel 2020-03-16 21:13:00 UTC 
+++ This bug was initially created as a clone of Bug #1812583 +++

Our default install is 3x4core masters, 3x2core workers. In an e2e run which is a representative workload for customers of medium to large scale clusters, we use

6.6 cores on average across masters (out of 12)
4.06 cores on average across workers (out of 6)

However, our default requests and limits for the pods components on this cluster are:

11.7 cores for the masters (out of 12)
8.04 cores for the workers (out of 6)

As a result, our default cluster cannot correctly install (runs out of requested CPU) but runs fine, which is a 4.4 release blocker.

---

CPU is compressible, and in general we don't set requests based on how much CPU you use, we establish ratios between components on the same roles.  For instance, etcd and kube-apiserver should receive the most CPU on the masters because if they get starved, all other components suffer.

The other master components should request CPU based on a formula like:

etcd request * (component usage / etcd usage)

The worker components that run on all nodes should follow a similar rule, based around either kubelet or openshift-sdn.  However, large components on the nodes may need a tweak factor:

sdn request * (component usage / sdn usage)


---

Request by namespace and role:

{namespace="openshift-monitoring",role="worker"}	4.009765625
{namespace="openshift-etcd",role="master"}	2.5810546875
{namespace="openshift-sdn",role="master"}	1.869140625
{namespace="openshift-sdn",role="worker"}	1.8046875
{namespace="openshift-kube-controller-manager",role="master"}	1.3212890625
{namespace="openshift-kube-apiserver",role="master"}	1.0810546875
{namespace="openshift-apiserver",role="master"}	0.90234375
{namespace="openshift-monitoring",role="master"}	0.736328125
{namespace="openshift-dns",role="worker"}	0.662109375
{namespace="openshift-dns",role="master"}	0.662109375
{namespace="openshift-controller-manager",role="master"}	0.603515625
{namespace="openshift-machine-config-operator",role="master"}	0.509765625
{namespace="openshift-image-registry",role="worker"}	0.466796875
{namespace="openshift-ingress",role="worker"}	0.40234375
{namespace="openshift-machine-config-operator",role="worker"}	0.240234375
{namespace="openshift-kube-storage-version-migrator",role="worker"}	0.201171875
{namespace="openshift-machine-api",role="master"}	0.181640625
{namespace="openshift-multus",role="master"}	0.12890625
{namespace="kube-system",role="master"}	0.123046875
{namespace="openshift-image-registry",role="master"}	0.10546875
{namespace="openshift-marketplace",role="worker"}	0.0859375
{namespace="openshift-console",role="master"}	0.0859375
{namespace="openshift-operator-lifecycle-manager",role="master"}	0.0859375
{namespace="openshift-cluster-node-tuning-operator",role="master"}	0.0859375
{namespace="openshift-kube-scheduler",role="master"}	0.0703125
{namespace="openshift-cluster-node-tuning-operator",role="worker"}	0.064453125
{namespace="openshift-multus",role="worker"}	0.064453125
{namespace="openshift-authentication",role="master"}	0.04296875
{namespace="openshift-dns-operator",role="master"}	0.041015625
{namespace="openshift-cluster-machine-approver",role="master"}	0.041015625
{namespace="openshift-cluster-version",role="master"}	0.041015625
{namespace="openshift-ingress-operator",role="master"}	0.041015625
{namespace="openshift-cluster-samples-operator",role="master"}	0.041015625
{namespace="openshift-network-operator",role="master"}	0.021484375
{namespace="openshift-kube-storage-version-migrator-operator",role="master"}	0.021484375
{namespace="openshift-service-ca-operator",role="master"}	0.021484375
{namespace="openshift-insights",role="master"}	0.021484375
{namespace="openshift-csi-snapshot-controller-operator",role="worker"}	0.021484375
{namespace="openshift-kube-controller-manager-operator",role="master"}	0.021484375
{namespace="openshift-authentication-operator",role="master"}	0.021484375
{namespace="openshift-cloud-credential-operator",role="master"}	0.021484375
{namespace="openshift-etcd-operator",role="master"}	0.021484375
{namespace="openshift-console-operator",role="master"}	0.021484375
{namespace="openshift-kube-apiserver-operator",role="master"}	0.021484375
{namespace="openshift-cluster-storage-operator",role="master"}	0.021484375
{namespace="openshift-controller-manager-operator",role="master"}	0.021484375
{namespace="openshift-service-catalog-controller-manager-operator",role="master"}	0.021484375
{namespace="openshift-service-ca",role="master"}	0.021484375
{namespace="openshift-marketplace",role="master"}	0.021484375
{namespace="openshift-csi-snapshot-controller",role="worker"}	0.021484375
{namespace="openshift-apiserver-operator",role="master"}	0.021484375
{namespace="openshift-kube-scheduler-operator",role="master"}	0.005859375
{namespace="openshift-service-catalog-apiserver-operator",role="master"}	0.005859375

Usage by namespace and role

{namespace="openshift-kube-apiserver",role="master"}	2.1619876333846624
{namespace="openshift-etcd",role="master"}	1.6578208562491923
{namespace="openshift-sdn",role="worker"}	0.7902127474534021
{namespace="openshift-apiserver",role="master"}	0.6654059737563104
{namespace="openshift-sdn",role="master"}	0.45042689393591295
{namespace="openshift-monitoring",role="worker"}	0.3504143483072635
{namespace="openshift-kube-controller-manager",role="master"}	0.20461836474549036
{namespace="openshift-etcd-operator",role="master"}	0.13743432789929938
{namespace="openshift-operator-lifecycle-manager",role="master"}	0.0884434181260195
{namespace="openshift-must-gather-kkjhl",role="master"}	0.08221540997755557
{namespace="openshift-monitoring",role="master"}	0.06295023150044035
{namespace="openshift-machine-config-operator",role="master"}	0.03892652122963432
{namespace="openshift-controller-manager",role="master"}	0.03838177104774991
{namespace="openshift-kube-apiserver-operator",role="master"}	0.03564176578869584
{namespace="openshift-ingress",role="worker"}	0.034161760609146746
{namespace="openshift-kube-scheduler",role="master"}	0.028997043612977058
{namespace="openshift-multus",role="master"}	0.028357733575176187
{namespace="openshift-cloud-credential-operator",role="master"}	0.02382491640890165
{namespace="openshift-kube-scheduler-operator",role="master"}	0.019677113395487122
{namespace="openshift-service-ca",role="master"}	0.019193497863481977
{namespace="openshift-kube-controller-manager-operator",role="master"}	0.017732005265559132
{namespace="openshift-marketplace",role="master"}	0.01595692508887505
{namespace="openshift-dns",role="worker"}	0.014646411228389108
{namespace="openshift-apiserver-operator",role="master"}	0.014566308284011975
{namespace="openshift-dns",role="master"}	0.013228066245094485
{namespace="openshift-image-registry",role="worker"}	0.013138524957335649
{namespace="openshift-marketplace",role="worker"}	0.012014092321005393
{namespace="openshift-console",role="master"}	0.007421924226788351
{namespace="openshift-image-registry",role="master"}	0.0071860119119124865
{namespace="openshift-authentication-operator",role="master"}	0.007069369592108443
{namespace="openshift-cluster-version",role="master"}	0.006795354406059571
{namespace="openshift-machine-config-operator",role="worker"}	0.006361723323325576
{namespace="openshift-authentication",role="master"}	0.006188662943334761
{namespace="openshift-machine-api",role="master"}	0.0052930512518087145
{namespace="openshift-network-operator",role="master"}	0.005136690421827466
{namespace="openshift-console-operator",role="master"}	0.004844998002650943
{namespace="openshift-controller-manager-operator",role="master"}	0.0045221224014901865
{namespace="openshift-multus",role="worker"}	0.0042029771037126965
{namespace="openshift-cluster-storage-operator",role="master"}	0.0038974091762590986
{namespace="openshift-service-ca-operator",role="master"}	0.0032243799219438558
{namespace="openshift-service-catalog-apiserver-operator",role="master"}	0.003034163254101611
{namespace="openshift-csi-snapshot-controller-operator",role="worker"}	0.002924888169013756
{namespace="openshift-service-catalog-controller-manager-operator",role="master"}	0.002260159790181888
{namespace="openshift-insights",role="master"}	0.002075358385262947
{namespace="openshift-cluster-samples-operator",role="master"}	0.002067038536157853
{namespace="openshift-cluster-node-tuning-operator",role="master"}	0.0019490981770514527
{namespace="openshift-kube-storage-version-migrator-operator",role="master"}	0.0018690529355203974
{namespace="kube-system",role="master"}	0.0013121596008085876
{namespace="openshift-ingress-operator",role="master"}	0.0012782041893306024
{namespace="openshift-csi-snapshot-controller",role="worker"}	0.0012313933297429486
{namespace="openshift-cluster-node-tuning-operator",role="worker"}	0.0009219469658359441
{namespace="openshift-cluster-machine-approver",role="master"}	0.0006950561574913539
{namespace="openshift-dns-operator",role="master"}	0.0006673132853684085
{namespace="openshift-kube-storage-version-migrator",role="worker"}	0.00018306567985376928

Request - usage

{namespace="openshift-monitoring",role="worker"}	3.6593512766927363
{namespace="openshift-sdn",role="master"}	1.418713731064087
{namespace="openshift-kube-controller-manager",role="master"}	1.1166706977545096
{namespace="openshift-sdn",role="worker"}	1.0144747525465978
{namespace="openshift-etcd",role="master"}	0.9232338312508075
{namespace="openshift-monitoring",role="master"}	0.6733778934995597
{namespace="openshift-dns",role="master"}	0.6488813087549055
{namespace="openshift-dns",role="worker"}	0.6474629637716109
{namespace="openshift-controller-manager",role="master"}	0.5651338539522501
{namespace="openshift-machine-config-operator",role="master"}	0.4708391037703657
{namespace="openshift-image-registry",role="worker"}	0.45365835004266436
{namespace="openshift-ingress",role="worker"}	0.3681819893908532
{namespace="openshift-apiserver",role="master"}	0.2369377762436895
{namespace="openshift-machine-config-operator",role="worker"}	0.23387265167667443
{namespace="openshift-kube-storage-version-migrator",role="worker"}	0.20098880932014623
{namespace="openshift-machine-api",role="master"}	0.1763475737481913
{namespace="kube-system",role="master"}	0.12173471539919141
{namespace="openshift-multus",role="master"}	0.10054851642482382
{namespace="openshift-image-registry",role="master"}	0.09828273808808752
{namespace="openshift-cluster-node-tuning-operator",role="master"}	0.08398840182294855
{namespace="openshift-console",role="master"}	0.07851557577321165
{namespace="openshift-marketplace",role="worker"}	0.07392340767899461
{namespace="openshift-cluster-node-tuning-operator",role="worker"}	0.06353117803416405
{namespace="openshift-multus",role="worker"}	0.060250147896287305
{namespace="openshift-kube-scheduler",role="master"}	0.04131545638702294
{namespace="openshift-dns-operator",role="master"}	0.040348311714631595
{namespace="openshift-cluster-machine-approver",role="master"}	0.04032056884250865
{namespace="openshift-ingress-operator",role="master"}	0.039737420810669395
{namespace="openshift-cluster-samples-operator",role="master"}	0.038948586463842146
{namespace="openshift-authentication",role="master"}	0.03678008705666524
{namespace="openshift-cluster-version",role="master"}	0.03422027059394043
{namespace="openshift-csi-snapshot-controller",role="worker"}	0.02025298167025705
{namespace="openshift-kube-storage-version-migrator-operator",role="master"}	0.019615322064479603
{namespace="openshift-insights",role="master"}	0.019409016614737054
{namespace="openshift-service-catalog-controller-manager-operator",role="master"}	0.01922421520981811
{namespace="openshift-csi-snapshot-controller-operator",role="worker"}	0.018559486830986245
{namespace="openshift-service-ca-operator",role="master"}	0.018259995078056146
{namespace="openshift-cluster-storage-operator",role="master"}	0.0175869658237409
{namespace="openshift-controller-manager-operator",role="master"}	0.016962252598509815
{namespace="openshift-console-operator",role="master"}	0.016639376997349055
{namespace="openshift-network-operator",role="master"}	0.016347684578172532
{namespace="openshift-authentication-operator",role="master"}	0.014415005407891557
{namespace="openshift-apiserver-operator",role="master"}	0.006918066715988025
{namespace="openshift-marketplace",role="master"}	0.00552744991112495
{namespace="openshift-kube-controller-manager-operator",role="master"}	0.0037523697344408677
{namespace="openshift-service-catalog-apiserver-operator",role="master"}	0.002825211745898389
{namespace="openshift-service-ca",role="master"}	0.0022908771365180228
{namespace="openshift-cloud-credential-operator",role="master"}	-0.00234054140890165
{namespace="openshift-operator-lifecycle-manager",role="master"}	-0.0025059181260194963
{namespace="openshift-kube-scheduler-operator",role="master"}	-0.013817738395487122
{namespace="openshift-kube-apiserver-operator",role="master"}	-0.014157390788695837
{namespace="openshift-etcd-operator",role="master"}	-0.11594995289929938
{namespace="openshift-kube-apiserver",role="master"}	-1.0809329458846624

We will update this shortly with the ratios that everyone should use.

--- Additional comment from Clayton Coleman on 2020-03-11 16:29:49 UTC ---

To gather these, grab an e2e run prometheus, find the time that the e2e tests mostly stopped, set that as the current time in your promecius or local prometheus instance in the graph selector, and run the queries:

(note the role stuff only works on GCP and Azure because AWS naming sucks)

Requests per namespace: sort_desc(sum by (namespace,role) ((max without (id,endpoint,image,job,metrics_path,instance,name,service) (label_replace(label_replace(container_spec_cpu_shares{pod!="",namespace!~"e2e.*"}, "role", "master", "node", ".*-m-.*"), "role", "worker", "node", ".*-w-.*")) / 1024)) > 0)

Usage per namespace (measured from the last 15m): sort_desc(sum by (namespace,role) ((max without (id,endpoint,image,job,metrics_path,instance,name,service) (label_replace(label_replace(rate(container_cpu_usage_seconds_total{pod!="",container="",namespace!~"e2e.*"}[15m]), "role", "master", "node", ".*-m-.*"), "role", "worker", "node", ".*-w-.*")))) > 0)

Difference between them: sort_desc(sort_desc(sum by (namespace,role) ((max without (id,endpoint,image,job,metrics_path,instance,name,service) (label_replace(label_replace(container_spec_cpu_shares{pod!="",namespace!~"e2e.*"}, "role", "master", "node", ".*-m-.*"), "role", "worker", "node", ".*-w-.*")) / 1024))) - sort_desc(sum by (namespace,role) ((max without (id,endpoint,image,job,metrics_path,instance,name,service) (label_replace(label_replace(rate(container_cpu_usage_seconds_total{pod!="",container="",namespace!~"e2e.*"}[15m]), "role", "master", "node", ".*-m-.*"), "role", "worker", "node", ".*-w-.*"))))))

Actual usage: sum by (role) (label_replace(label_replace(rate(container_cpu_usage_seconds_total{id="/"}[15m]), "role", "master", "node", ".*-m-.*"), "role", "worker", "node", ".*-w-.*"))

--- Additional comment from Clayton Coleman on 2020-03-11 17:53:29 UTC ---

After some basic data analysis, it's reasonable to say that out of the 6.6 cores in use, the fraction used by the key components on masters is:

25% etcd 25%
33% kube-apiserver
10% openshift-apiserver
5%  kcm

That's 73% of total usage.  I would expect the requests to be roughly proportional to these percentages out of our arbitrary floor.  I think 3 cores requested is a reasonable base master spec for idle, in which case the requests would be:

330m etcd
250m kube-apiserver
 10m openshift-apiserver
  5m kcm

And then the remaining components should have requests that consume no more than 270m, divied up fairly based on their average use.  A good default is going to be 5m for an operator that doesn't serve traffic or answer lots of queries.

This would end up with us on masters having roughly 1 core set aside for core workload, and we would only schedule flex workloads on masters down to that single core.  In very large masters these numbers might have to flex upwards, but we can't solve that component by component.

--- Additional comment from Clayton Coleman on 2020-03-11 18:40:20 UTC ---

Working on draft recommendations.

--- Additional comment from Clayton Coleman on 2020-03-11 22:31:34 UTC ---

Here is the draft recommendation that child bugs should follow:

Rules for component teams:

1. Determine your average CPU usage from the list above (breaking down any components that are split across namespace)

2. Pods that run on a master that are not on all nodes (i.e. exclude dns, openshift-sdn, mcd) should have a request that is proportional to your CPU usage relative to kube-apiserver.

kube-apiserver is allocated 33% of all resources (2.16 cores out of 6.6 cores). Calculate your CPU usage relative to kube-apiserver, and then multiply 330m by your fraction of kube-apiserver use (i.e. kube-scheduler uses 0.028 cores, so 0.028 * 330m = 9.24m)

Special cases:
* Certain infra components will be assigned slightly higher requests (kcm, scheduler, ocm given known problems if they fall behind)
* Leader elected components should set their request to their expected usage across all pods, even though they will be over provisioned (kcm should request Xm on all nodes, not Xm/3)
* No component should be lower than 5m per pod, but it must be set to 5m

3. Pods that run on a worker should be proportional to actual CPU use on our minimum e2e run. openshift-sdn uses 790m on workers, so between ovs and sdn per node there should be 790m/3 ~ 250m of CPU allocation between ovs and sdn pods.

4. Large infra components like monitoring should set their request proportional to openshift-sdn namespace

openshift-sdn is allocated 750m and uses 790m.  Prometheus uses 350m and requests 4 on workers.  Because prometheus has a significant vertical scaling component, it should probably be close to openshift-sdn in terms of requests, and if it needs more active resource growth the operator should manage that.  Node-exporter should be set relative to openshift-sdn

--- Additional comment from Clayton Coleman on 2020-03-11 22:41:27 UTC ---

Commit message recommendation:

    Normalize CPU requests on masters

    The {x} uses approximately {percent}% of master CPU in a
    reasonable medium sized workload. Given a 1 core per master baseline
    (since CPU is compressible and shared), assign the kube-apiserver
    roughly {desired_percent}% of that core on each master.

PRs will be opened tracking each component in the list above (core control plane) and then the worst offenders in the list will be updated.

--- Additional comment from Clayton Coleman on 2020-03-11 22:56:24 UTC ---

Allocating 10% for the kube-controller-manager (1/3 of kube-apiserver).

--- Additional comment from Clayton Coleman on 2020-03-11 23:05:56 UTC ---

openshift-apiserver is 10% (1/3 of kube-apiserver)
etcd is 25% (~70-80% of kube-apiserver)

--- Additional comment from Clayton Coleman on 2020-03-11 23:25:33 UTC ---

On the control plane the remaining components are expected to allocate (100 - 33 - 25 - 10 - 10 = ) 220m between them.

The most aggressive components will be targeted first - monitoring is the highest requester.

On the nodes, SDN is reasonably sized right now, and OVS is using about half of its 200m request.  SDN is slightly more expensive (why?) so we should tune OVS down to 100m for now.
Comment 2 Scott Dodson 2020-03-17 13:20:06 UTC 

*** This bug has been marked as a duplicate of bug 1812709 ***
Comment 3 Red Hat Bugzilla 2023-09-14 05:54:25 UTC 
The needinfo request[s] on this closed bug have been removed as they have been unresolved for 1000 days