Provider Hardware Requirements and Recommendations
Software Recommendation
Providers have been tested on Ubuntu 22.04 with the default Linux kernel. Your experience may vary if installation is attempted using a different Linux distro/kernel.
Kubernetes Control Plane (Master) Node Requirements
Minimum Specs
- 4 CPU
- 8 GB RAM
- 50 GB disk
Recommended Specs
- 8 CPU
- 16 GB RAM
- 100 GB disk
Kubernetes Worker Node Requirements
Minimum Specs
- 4 CPU
- 8 GB RAM
- 100 GB disk
Recommendations
- The more resources you have, the better, especially if your goal is to maximize the number of concurrent deployments.
- It is particularly important to note that the worker node needs to have as much CPU capacity as possible. For instance, if the node has 8 CPUs, 100 GB of RAM, and 2 TB of disk space, the CPU would likely become the bottleneck.
- Since users tend to allocate at least 1 CPU per deployment, the server could host a maximum of 8 deployments, but realistically it will likely support about 6 deployments, as approximately 2 CPUs will be reserved for Kubernetes system components.
Creating a Provider
To create a provider, you can either create a single server provider or have a small server for the control plane and other larger servers for the worker nodes. It is recommended to have a small control plane server for the Spheron provider and add larger servers as workers so that all the compute or GPU resources are used for user deployments.
It is important to note that the recommended minimum number of hosts is 4 for a production Provider Kubernetes cluster. This setup allows:
- 3 hosts serving as redundant control plane (master)/etcd instances.
- 1 host to serve as a Kubernetes worker node to host provider leases.
Real-Life Use Case Recommendations
- In a typical production Kubernetes cluster, 3 redundant control plane nodes are recommended. However, when the control plane node is easily recoverable, a single control instance for providers can be sufficient.
- The number of control plane nodes should always be odd to ensure the cluster can reach consensus.
- It is recommended to run a single worker node per physical server, as CPU is generally the primary resource bottleneck. Using a single worker node enables larger workloads to be deployed on your provider.
NOTE: For initial testing, we recommend running a single server provider with a GPU to test the entire process. After successful testing, you can then move to adding multiple nodes with better hardware specifications.
etcd Hardware Recommendations
Use this guide (opens in a new tab) to ensure Kubernetes control plane nodes meet the recommendations for hosting an etcd database.