Skip to content

Using the Karpenter Autoscaler and ARM Processors: Benefits, Optimization, and Analytics

Written by:

Igor Gorovyy
DevOps Engineer Lead & Senior Solutions Architect
LinkedIn

Max Ivashchenko
Senior Solutions Architect at Amazon Web Services (AWS)
LinkedIn

Introduction

Modern cloud solutions demand maximum efficiency, performance, and cost savings. The combination of ARM processors and the Karpenter autoscaler in Amazon EKS offers significant advantages for developers and DevOps teams. This article explores how this combination works, its key benefits, and the analytics metrics that can enhance infrastructure efficiency.

What is Karpenter?

Karpenter is an intelligent Kubernetes autoscaler that dynamically scales clusters based on workload needs. Its key features include:

  • Dynamic node provisioning and deletion based on workload demand.
  • Cost optimization through the use of reserved and spot instances.
  • Support for various instance types and processors, including ARM-based Graviton.

Why ARM Processors?

ARM processors, such as AWS Graviton, provide high performance and energy efficiency at a lower cost than x86 processors. Key advantages include:

  • Cost Savings: Up to 40% reduction in instance costs.
  • Energy Efficiency: Reduced energy consumption.
  • Performance: Optimized for modern workloads, including containers, microservices, and ML.

Setting Up Karpenter for ARM

1. Choosing Suitable Instances

  • For ARM: Use instance types such as t4g, m6g, c7g, and other Graviton-based options.
  • Configure Karpenter provisioners to prioritize these instance types.

2. Tags and Configuration

Add parameters in your provisioners to target ARM processors.

Example:

```yaml architecture: arm64 instanceTypes: - m6g.large - c7g.medium

3. Spot Instance Optimization

Karpenter can leverage spot instances for additional savings. Ensure SQS is configured to monitor the status of spot instances.

Analytics Metrics

To effectively use the autoscaler and ARM processors, monitor metrics such as:

Resource Utilization:
    CPU and RAM usage on ARM instances.
    Node usage percentage (under- or overutilization).

Scaling Time:
    Average time to provision and delete nodes.
    Impact on workload performance during scaling.

Cost Savings:
    Cost comparison of ARM vs. x86 instances.
    Spot instance utilization percentage.

Workload Dynamics:
    CPU/Memory requests by pods.
    Traffic and workload changes on the cluster.

Error Rate:
    Image pulling errors.
    Issues with spot instances (e.g., premature termination).

Examples

Example 1: Configuring Karpenter for ARM Processors

Task: Configure Karpenter to create ARM-based Graviton nodes optimized for web service workloads.

Provisioner Configuration:

```yaml apiVersion: karpenter.sh/v1alpha5 kind: Provisioner metadata: name: arm-optimized spec: provider: instanceProfile: "KarpenterNodeInstanceProfile" requirements: - key: "karpenter.k8s.aws/instance-type" operator: In values: - "m6g.large" - "m6g.xlarge" - "c7g.medium" - key: "kubernetes.io/arch" operator: In values: - "arm64" limits: resources: cpu: "200" memory: "400Gi" ttlSecondsAfterEmpty: 30

Comparing Karpenter with Cluster Autoscaler

Feature Cluster Autoscaler Karpenter
Scaling Method Adds/removes nodes based on predefined Node Groups. Dynamically creates nodes based on pod needs.
Instance Selection Limited to predefined Node Groups. Supports all instance types, including ARM (Graviton).
Cost Optimization Limited to predefined instance types. Optimized with ARM and spot instances.
Architecture Support No specific ARM support. Seamless integration with ARM (arm64).
Scaling Speed Slower (~2-3 minutes). Faster (~1-2 minutes).
Resource Utilization May lead to underutilized resources. Precise node scaling minimizes waste.