diff options
| author | Mateusz Bularz <60339703+M4itee@users.noreply.github.com> | 2023-11-27 10:45:18 +0100 |
|---|---|---|
| committer | Mateusz Bularz <60339703+M4itee@users.noreply.github.com> | 2023-11-27 10:45:18 +0100 |
| commit | 54d59a9a3a5127c5eee909ee311fc9bfba054486 (patch) | |
| tree | 009b2bb0c8259b6973f0d99da61763f95913d51f | |
| parent | 29693397a528e6f9355bbc59c5e3898fbf06c633 (diff) | |
adding info about hardware requirements
| -rw-r--r-- | docs/netdata-cloud-onprem/getting-started.md | 29 |
1 files changed, 29 insertions, 0 deletions
diff --git a/docs/netdata-cloud-onprem/getting-started.md b/docs/netdata-cloud-onprem/getting-started.md index 482214e24d..dd0c6d5bd2 100644 --- a/docs/netdata-cloud-onprem/getting-started.md +++ b/docs/netdata-cloud-onprem/getting-started.md @@ -22,6 +22,35 @@ Helm charts are designed for kubernetes to run as the local equivalent of the ne - Default storage class configured and working (Persistent volumes based on SSDs are preferred) `*` - available in dependencies helm chart for PoC applications. +#### Hardware requirements: +##### How we tested it: +- A number of VMs on the AWS EC2, size of the instance was c6a.32xlarge (128CPUs / 256GiB memory). +- Host system - Ubuntu 22.04. +- Each VM hosts 200 Agent nodes as docker containers. +- Agents are connected DIRECTLY to cloud (no Parent-Child relationships). This is the worst option for the cloud. +- Cloud hosted on 1 kubernetes node c6a.8xlarge (32CPUs / 64GiB memory). +- Dependencies were also installed on the same node. +- Maximum connected nodes was ~2000. + +##### Results +There was no point in trying to connect more nodes as we are covering the PoC purposes. +- In a peak connection phase - All nodes startup were triggered in ~15 minues: + - Up to 60% (20 cores) CPU usage of the kubernetes node. Top usage came from: + - Ingress controller (we used haproxy ingress controller) + - Postgres + - Pulsar + - EMQX + Combined they were responsible for ~30-35% of CPU usage of the node. +- When all nodes connected and synchronized their state CPU usage floated between 30% and 40% - depending on what we did on the cloud (browsing different). Here top offenders were: + - Pulsar + - Postgres + Combined they were responsible for ~15-20% of CPU usage of the node. +- Memory usage - 45GiB in a peak. Most of it (~20GiB) was consumed by: + - Postgres + - Elastic + - Pulsar + +For a comparison - Netdata Cloud On-prem installation with just 100 nodes connected, without dependencies is going to consume ~2CPUs and ~2GiB of memory (REAL usage, not requests on a Kubernetes). ## Pulling the helm chart Helm chart for the Netdata Cloud On-Prem installation on Kubernetes is available at ECR registry. |