Skip to main content

Provisioning AWS EFS

Overview​

This guide explains how to provision an AWS EFS-backed shared storage for llm-d using the EFS CSI driver.

Prerequisites​

  • Ensure your cluster infrastructure is sufficient to deploy high scale inference.
  • An AWS account.
  • An EKS cluster.
  • An existing EFS filesystem.
  • Create a namespace for installation.
export NAMESPACE=llm-d-storage # or any other namespace (shorter names recommended)
kubectl create namespace ${NAMESPACE}

Cluster setup for provisioning AWS EFS (EKS cluster)​

Install EFS CSI Driver​

Follow the official AWS guide to install the EFS CSI driver:

https://docs.aws.amazon.com/eks/latest/userguide/efs-csi.html

Ensure the driver is running in your cluster:

kubectl get pods -n kube-system | grep efs

Create an EFS File System​

Create an EFS filesystem in the same VPC as your EKS cluster and note the fileSystemId.

You will need this ID when configuring the StorageClass.

Provisioning​

EKS​

1. Create a StorageClass for EFS:

cd guides/tiered-prefix-cache/storage/manifests/backends/aws
kubectl apply -f ./storage_class.yaml -n ${NAMESPACE}

Update the fileSystemId in storage_class.yaml before applying.

2. Create a PVC:

apiVersion: v1
kind: PersistentVolumeClaim
metadata:
  name: llm-d-kv-cache-storage
spec:
  accessModes:
  - ReadWriteMany
  resources:
    requests:
      storage: 1Ti
  storageClassName: efs-sc

Check the PVC​

kubectl get pvc -n ${NAMESPACE}

Output should show the PVC as Bound:

NAME                        STATUS   VOLUME   CAPACITY   ACCESS MODES   STORAGECLASS   AGE
llm-d-kv-cache-storage      Bound    ...      ...        RWX            efs-sc         1m

Performance Recommendations​

For best performance with AWS EFS:

  • Use Max I/O performance mode for higher throughput
  • Prefer Provisioned Throughput for consistent performance under heavy workloads
  • Mount EFS in the same VPC and availability zone as your EKS nodes
  • Use multiple vLLM replicas to parallelize I/O
  • Monitor throughput and latency using AWS CloudWatch metrics

Note: EFS has higher latency than local SSD or Lustre, and is best suited for shared cache reuse rather than ultra-low latency workloads.

Cleanup​

kubectl delete -f ./storage_class.yaml -n ${NAMESPACE}