63-Node EKS Cluster running on a Single Instance with Firecracker

63-Node EKS Cluster running on a Single Instance with Firecracker

This blog post is a part of a series of posts devoted to Firecracker automation. Currently it consists of the following posts:

Elaborating on the original code created to automate the creation of Firecracker VMs, this post will show how we can create an EKS-compatible (as in Amazon’s EKS) cluster, with multiple nodes, where each node will be a Virtual Machine, all running on a single host, thanks to Firecracker.

All source code is available in this GitLab repository.

Creating an EKS Cluster out of Firecracker VMs

EKS? Isn’t that an AWS managed Kubernetes service? It certainly is, but it also is open source recently announced and contributed by Amazon Web Services. The latter form is more precisely called EKS Distro (EKS-D), and its source code is available on this Github repository.

In particular, Ubuntu’s snap-based EKS-D compatible distribution will be used, since it is quite easy to install and operate on Ubuntu VMs. It is interesting that this distribution is based on microk8s, and is labeled as “EKS-compatible Kubernetes”. “Compatibility” in this context appears to be that is kind of a microk8s “dressed as” EKS, as it includes the eks binary, which works as expected elsewhere. But because it is based on microk8s, it is probably suitable only for those cases where microk8s is.

With the code developed as part of the previous post, a configurable number of Firecracker VMs can be created automatically, with password-less SSH configured with user fc, which can also password-less sudo. Given this, and that we need to install and configure the instances to install Ubuntu’s EKS-D, Ansible was a perfect candidate for this automation.

All of the Ansible code can be found in the ansible/ folder. I’m not an Ansible expert, so code is subject to many improvements (feel free to submit MRs!). Code should be quite self-explanatory –except for the inventory part, which required some additional shell scripting that can be found in the 05-install_eks_via_ansible.sh file.

There are two problems that we need to solve with the shell script:

  • Generate a dynamic inventory. Basically, the number of VMs is based on the NUMBER_VMS variable defined in the variables file. The first node will be considered the EKS master, and all remaining nodes workers.

  • After installing the eks snap, each node will be a separate “master”. To form a single cluster, we need to run eks add-node command in the master for every node, and an equivalent eks join in the nodes. Both commands use a token, which can be auto-generated (not convenient for this use case) or provided. The shell script generates these tokens and provide them as variables: a token for each node, and a list of all tokens to the master node.

You may run this code in your own environment. Review the README.md file and make sure firecracker is installed and properly configured, as described in the first post. 4GB and 2 cores seems to be a reasonable minimum of resources needed per VM to run a successful EKS-D compatible cluster. Adjust variables file to your environment.

Creating a 63-Node EKS Cluster on a r5d.metal instance

Wouldn’t it be cool to create a large EKS cluster, running on many VMs… all within a single “big” host? That’s the experiment I run. I chose a r5d.metal instance (96 cores, 768 GB RAM, 3.6 TB NVMe local SSD), as the best combination of high core and RAM count and local storage (as we will be creating the “volumes” for the VMs, and this should be as fast as possible). Then fired 100 Firecracker VMs and created an EKS cluster…

I set the number of Ansible forks to 32, to have more parallelism. However, it looks like even with low parallelism there are errors with the eks join operation caused by limits on the maximum number of requests to the kubeapi-server, so this part of the Ansible code is serialized. VMs configured with 2 core and 4 GB of RAM. The code worked. Scripts, ansible, all worked. A bit after 4 hours, a 100-node EKS cluster created within a single instance was running:

Insert benchmark comparison

However, load was at more than 200, and while memory usage was acceptable (2/3rds of the system RAM), cluster was unusable. But more importantly, I found an interesting fact: kubectl get nodes would only return 63 nodes, despite the 99 node join operations were returned as successful. I didn’t dig very much, but there seems to be a limitation here (internal microk8s limitation? Not sure anyway microk8s has been extensively tested with this high number of cluster instances…).

So I decided to repeat the operation, but with “only” 63 VMs (after all, kubectl get nodes won’t return more…) and bump the VM specs to 4 core (the master node VM was also loaded over 100% itself) and 16 GB of RAM (16 GB * 63 is more than the system’s available RAM, but Firecracker allows for both memory and CPU overcommitment, as long as not all processes use all the allocated RAM). This worked better, cluster was successfully created and kubectl get nodes reported this time all the nodes. All in about 3h, now with an acceptable CPU load:

Insert benchmark comparison

Insert benchmark comparison

ubuntu@ip-172-31-93-220:~$ kubectl get nodes
id1394205090   Ready    <none>   116m    v1.18.9-eks-1-18-1
id0890521735   Ready    <none>   93m     v1.18.9-eks-1-18-1
id1865400204   Ready    <none>   85m     v1.18.9-eks-1-18-1
id3176525421   Ready    <none>   79m     v1.18.9-eks-1-18-1
id0289819564   Ready    <none>   97m     v1.18.9-eks-1-18-1
id2364814864   Ready    <none>   101m    v1.18.9-eks-1-18-1
id1576708969   Ready    <none>   62m     v1.18.9-eks-1-18-1
id0839217590   Ready    <none>   95m     v1.18.9-eks-1-18-1
id2385921825   Ready    <none>   122m    v1.18.9-eks-1-18-1
id2162009048   Ready    <none>   99m     v1.18.9-eks-1-18-1
id0032931251   Ready    <none>   94m     v1.18.9-eks-1-18-1
id1397218544   Ready    <none>   83m     v1.18.9-eks-1-18-1
id2509806641   Ready    <none>   72m     v1.18.9-eks-1-18-1
id1391026381   Ready    <none>   56m     v1.18.9-eks-1-18-1
id3180021860   Ready    <none>   132m    v1.18.9-eks-1-18-1
id2766207659   Ready    <none>   130m    v1.18.9-eks-1-18-1
id2417208994   Ready    <none>   67m     v1.18.9-eks-1-18-1
id0037315342   Ready    <none>   126m    v1.18.9-eks-1-18-1
id2870908982   Ready    <none>   124m    v1.18.9-eks-1-18-1
id0303528979   Ready    <none>   96m     v1.18.9-eks-1-18-1
id2443620467   Ready    <none>   100m    v1.18.9-eks-1-18-1
id2671116621   Ready    <none>   89m     v1.18.9-eks-1-18-1
id1153919941   Ready    <none>   127m    v1.18.9-eks-1-18-1
id1618117867   Ready    <none>   114m    v1.18.9-eks-1-18-1
id1345427105   Ready    <none>   128m    v1.18.9-eks-1-18-1
id1578112374   Ready    <none>   133m    v1.18.9-eks-1-18-1
id2260431270   Ready    <none>   134m    v1.18.9-eks-1-18-1
id1192720052   Ready    <none>   131m    v1.18.9-eks-1-18-1
id0009731256   Ready    <none>   86m     v1.18.9-eks-1-18-1
id0772428726   Ready    <none>   112m    v1.18.9-eks-1-18-1
id2345631371   Ready    <none>   115m    v1.18.9-eks-1-18-1
id0866401320   Ready    <none>   90m     v1.18.9-eks-1-18-1
id2295817569   Ready    <none>   121m    v1.18.9-eks-1-18-1
id3195221619   Ready    <none>   105m    v1.18.9-eks-1-18-1
id3212514411   Ready    <none>   64m     v1.18.9-eks-1-18-1
id2530228284   Ready    <none>   58m     v1.18.9-eks-1-18-1
id1582731567   Ready    <none>   125m    v1.18.9-eks-1-18-1
id0234531685   Ready    <none>   104m    v1.18.9-eks-1-18-1
id1170407709   Ready    <none>   129m    v1.18.9-eks-1-18-1
id2377401525   Ready    <none>   129m    v1.18.9-eks-1-18-1
id2782930529   Ready    <none>   109m    v1.18.9-eks-1-18-1
id2316206444   Ready    <none>   88m     v1.18.9-eks-1-18-1
id0886017938   Ready    <none>   120m    v1.18.9-eks-1-18-1
id2886224137   Ready    <none>   135m    v1.18.9-eks-1-18-1
id2137806005   Ready    <none>   74m     v1.18.9-eks-1-18-1
id3252717927   Ready    <none>   136m    v1.18.9-eks-1-18-1
id1421103858   Ready    <none>   111m    v1.18.9-eks-1-18-1
id2048704797   Ready    <none>   102m    v1.18.9-eks-1-18-1
id1892014011   Ready    <none>   106m    v1.18.9-eks-1-18-1
id0719403446   Ready    <none>   61m     v1.18.9-eks-1-18-1
id1134505657   Ready    <none>   82m     v1.18.9-eks-1-18-1
id0023523299   Ready    <none>   118m    v1.18.9-eks-1-18-1
id2477116066   Ready    <none>   81m     v1.18.9-eks-1-18-1
id0469627718   Ready    <none>   77m     v1.18.9-eks-1-18-1
id1020429150   Ready    <none>   113m    v1.18.9-eks-1-18-1
id1471722684   Ready    <none>   91m     v1.18.9-eks-1-18-1
id1417322458   Ready    <none>   75m     v1.18.9-eks-1-18-1
id0107720352   Ready    <none>   69m     v1.18.9-eks-1-18-1
id2957917156   Ready    <none>   108m    v1.18.9-eks-1-18-1
id0573606644   Ready    <none>   3h54m   v1.18.9-eks-1-18-1
id1377114631   Ready    <none>   117m    v1.18.9-eks-1-18-1
id0468822693   Ready    <none>   119m    v1.18.9-eks-1-18-1
id0660532190   Ready    <none>   71m     v1.18.9-eks-1-18-1

Despite all this, the cluster was not really usable. Most operations took long to complete and some errored out. Probably, the microk8s on which this EKS-D is based is not designed for this purposes ;)

Final words

The 63- and 100-Node experiment was more of a funny exercise and a validation for the scripts and Ansible code. However, the code presented is quite useful specially for testing scenarios. I can create on my laptop a 3-node EKS cluster (2 core, 4 GB of RAM per node) in under 5 minutes, all with a single-line command. And destroy it all in seconds with another one-liner. This allows for quick testing of EKS-compatible Kubernetes clusters. All thanks to Firecracker and EKS-D, both open source componentes released by AWS. Thanks!