EKS – best instance type to use as worker nodes optimising for price per IP address
By Tom In DevOps
We are playing with the new Elastic Container Service for Kubernetes from Amazon (EKS) and have found it interesting that instead of using a overlay network like we have been used to in previous Kubernetes experiments, they are assigning IP addresses to each Pod via their Elastic Network Adaptors. The plugin they use is open-sourced (amazon-vpc-cni-k8s). I like the simplicity of the idea so that network packets aren’t needlessly wrapped in an additional layer but the problem is that there is a limited number of IPs per instance and therefore it restricts the number of Pods that can be run on each node.
If running average applications, the pod limit might not be a issue because the CPU or memory would be the bottleneck but if you are running small mircroservices, or many CronJobs, the limiting factor for a node is the number of IP addresses. I did some calculations to find the best AWS instance type in terms of cheapest per IP:
API Name Memory vCPUs Max IPs ENIs Cost Cost per IP t2.nano 0.5 GiB 1 vCPUs 4 2 0.0058 0.00145 t2.medium 4.0 GiB 2 vCPUs 18 3 0.0464 0.0025777778 t2.large 8.0 GiB 2 vCPUs 36 3 0.0928 0.0025777778 c5.large 4.0 GiB 2 vCPUs 30 3 0.085 0.0028333333 c5.xlarge 8.0 GiB 4 vCPUs 60 4 0.17 0.0028333333 c5.4xlarge 32 GiB 16 vCPU 240 8 0.68 0.0028333333 t2.small 2.0 GiB 1 vCPUs 8 2 0.023 0.002875 t2.micro 1.0 GiB 1 vCPUs 4 2 0.0116 0.0029 m5.large 8.0 GiB 2 vCPUs 30 3 0.096 0.0032 m5.xlarge 16 GiB 4 vCPUs 60 4 0.192 0.0032 m5.4xlarge 64 GiB 16 vCPUa 240 8 0.768 0.0032 c5d.large 4.0 GiB 2 vCPUs 30 3 0.096 0.0032 c5d.xlarge 8.0 GiB 4 vCPUs 60 4 0.192 0.0032 c5d.4xlarge 32 GiB 16 vCPUa 240 8 0.768 0.0032 c4.xlarge 7.5 GiB 4 vCPUs 60 4 0.199 0.0033166667 c4.4xlarge 30.0 GiB 16 vCPUa 240 8 0.796 0.0033166667 m4.xlarge 16.0 GiB 4 vCPUs 60 4 0.2 0.0033333333 m4.4xlarge 64.0 GiB 16 vCPUa 240 8 0.8 0.0033333333 c4.large 3.75 GiB 2 vCPUs 30 3 0.1 0.0033333333 c3.large 3.75 GiB 2 vCPUs 30 3 0.105 0.0035 c3.xlarge 7.5 GiB 4 vCPUs 60 4 0.21 0.0035 c3.4xlarge 30.0 GiB 16 vCPUs 240 8 0.84 0.0035 m5d.large 8.0 GiB 2 vCPUs 30 3 0.113 0.0037666667 m5d.xlarge 16.0 GiB 4 vCPUs 60 4 0.226 0.0037666667 m5d.4xlarge 64.0 GiB 16 vCPUs 240 8 0.904 0.0037666667 h1.4xlarge 64.0 GiB 16 vCPUs 240 8 0.936 0.0039 c5.18xlarge 144.0 GiB 72 vCPUs 750 15 3.06 0.00408 m2.xlarge 17.1 GiB 2 vCPUs 60 4 0.245 0.0040833333 m2.2xlarge 34.2 GiB 4 vCPUs 120 4 0.49 0.0040833333
This is for the Linux hourly on-demand price. I think we’re going to go with t2.large but making sure the unlimited option is set so they don’t run out of credits.
If you want this data yourself, my source was www.ec2instances.info – you can change the columns shown to relevant ones, then copy&paste the table into LibreOffice to calculate the price per IP and sort it.
21/Aug/18 – The t3 instance type was announced and it has the best cost per IP ratio.
No Comment