Source: examples/aws_efa

Using Elastic Fabric Adapter (EFA) on AWS with SkyPilot#

Elastic Fabric Adapter (EFA) is an AWS alternative to Nvidia infiniband that enables high levels of inter-node communications. It is specifically useful for distributed AI training and inference, which requires high network bandwidth across nodes.

Using EFA on HyperPod/EKS with SkyPilot#

TL;DR: enable EFA with SkyPilot#

You can enable EFA on AWS HyperPod/EKS clusters with an simple additional setting in your SkyPilot YAML:

config:
  kubernetes:
    pod_config:
      spec:
        containers:
        - resources:
            limits:
              vpc.amazonaws.com/efa: 4
            requests:
              vpc.amazonaws.com/efa: 4

Enable EFA with HyperPod/EKS#

  • On HyperPod (backed by EKS), EFA is enabled by default, and you don’t need to do anything.

  • On EKS, you may need to enable EFA with the official AWS documentation.

To check if EFA is enabled, run:

kubectl get nodes "-o=custom-columns=NAME:.metadata.name,INSTANCETYPE:.metadata.labels.node\.kubernetes\.io/instance-type,GPU:.status.allocatable.nvidia\.com/gpu,EFA:.status.allocatable.vpc\.amazonaws\.com/efa"

You can expect a sample output like:

NAME                           INSTANCETYPE      GPU   EFA
hyperpod-i-0beea7c849d1dc614   ml.p4d.24xlarge   8     4
hyperpod-i-0da69b9076c7ff6a4   ml.p4d.24xlarge   8     4
...

Access HyperPod and run distributed job with SkyPilot#

To access HyperPod and run distributed job with SkyPilot, see the SkyPilot HyperPod example.

Adding EFA configurations in SkyPilot YAML#

To enable EFA in SkyPilot YAML, you can specify the following section in the SkyPilot YAML:

config:
  kubernetes:
    pod_config:
      spec:
        containers:
        - resources:
            limits:
              vpc.amazonaws.com/efa: 4
            requests:
              vpc.amazonaws.com/efa: 4

This section is important for EFA integration:

  • config.kubernetes.pod_config: Provides Kubernetes-specific pod configuration

  • spec.containers[0].resources: Defines resource requirements

    • limits.vpc.amazonaws.com/efa: 4: Limits the Pod to use 4 EFA devices

    • requests.vpc.amazonaws.com/efa: 4: Requests 4 EFA devices for the Pod

The vpc.amazonaws.com/efa resource type is exposed by the AWS EFA device plugin in Kubernetes. To see how many EFA are available for each instance types that have EFA, see the Network cards list in the Amazon EC2 User Guide.

Check the following table for the GPU and EFA count mapping for AWS instance types:

Instance Type

GPU Type

#EFA

p4d.24xlarge

A100:8

4

p4de.24xlarge

A100:8

4

p5.48xlarge

H100:8

32

p5e.48xlarge

H200:8

32

p5en.48xlarge

H200:8

16

g5.8xlarge

A10G:1

1

g5.12xlarge

A10G:4

1

g5.16xlarge

A10G:1

1

g5.24xlarge

A10G:4

1

g5.48xlarge

A10G:8

1

g4dn.8xlarge

T4:1

1

g4dn.12xlarge

T4:4

1

g4dn.16xlarge

T4:1

1

g4dn.metal

T4:8

1

g6.8xlarge

L4:1

1

g6.12xlarge

L4:4

1

g6.16xlarge

L4:1

1

g6.24xlarge

L4:4

1

g6.48xlarge

L4:8

1

g6e.8xlarge

L40S:1

1

g6e.12xlarge

L40S:4

1

g6e.16xlarge

L40S:1

1

g6e.24xlarge

L40S:4

2

g6e.48xlarge

L40S:8

4

Update the EFA number in the nccl_efa.yaml for the GPUs you use.

Running NCCL test with EFA using SkyPilot#

Check the nccl_efa.yaml for the complete SkyPilot cluster yaml configurations.

The image_id provides the environment setup for NCCL (NVIDIA Collective Communications Library) and EFA (Elastic Fabric Adapter).

To run the NCCL test with EFA support:

sky launch -c efa nccl_efa.yaml

SkyPilot will:

  1. Schedule the job on a Kubernetes cluster with EFA-enabled nodes

  2. Launch Pods with the required EFA devices

  3. Execute the NCCL performance test with EFA networking

  4. Output performance metrics showing the benefits of EFA for distributed training

NOTE: We can turn off EFA with nccl_efa.yaml by passing an env:

sky launch -c efa --env USE_EFA=false nccl_efa.yaml

Benchmark results#

We compare the performance with and without EFA using NCCL test reports on the same HyperPod cluster (2x p4d.24xlarge, i.e. 2xA100:8).

The Speed-up column is calculated by busbw EFA (GB/s) / busbw Non-EFA (GB/s).

Message Size

busbw EFA (GB/s)

busbw Non-EFA (GB/s)

Speed-up

8 B

0

0

-

16 B

0

0

-

32 B

0

0

-

64 B

0

0

-

128 B

0

0

-

256 B

0

0

-

512 B

0.01

0.01

1 x

1 KB

0.01

0.01

1 x

2 KB

0.02

0.02

1 x

4 KB

0.04

0.05

0.8 x

8 KB

0.08

0.06

1.3 x

16 KB

0.14

0.06

2.3 x

32 KB

0.25

0.17

1.4 x

64 KB

0.49

0.23

2.1 x

128 KB

0.97

0.45

2.1 x

256 KB

1.86

0.68

2.7 x

512 KB

3.03

1.01

3 x

1 MB

4.61

1.65

2.8 x

2 MB

6.5

1.75

3.7 x

4 MB

8.91

2.39

3.7 x

8 MB

10.5

2.91

3.6 x

16 MB

19.03

3.22

5.9 x

32 MB

31.85

3.58

8.9 x

64 MB

44.37

3.85

11.5 x

128 MB

54.94

3.87

14.2 x

256 MB

65.46

3.94

16.6 x

512 MB

71.83

4.04

17.7 x

1 GB

75.34

4.08

18.4 x

2 GB

77.35

4.13

18.7 x

What stands out#

Range

Observation

≤ 256 KB

Virtually no difference — bandwidth is dominated by software/latency overhead and doesn’t reach the network’s limits.

512 KB – 16 MB

EFA gradually pulls ahead, hitting ~3–6 × by a few MB.

≥ 32 MB

The fabric really kicks in: ≥ 8 x at 32 MB, climbing to ~18 x for 1–2 GB messages. Non-EFA tops out around 4 GB/s, while EFA pushes ≈ 77 GB/s.

EFA provides much higher throughput than the traditional TCP transport. Enabling EFA could enhance the performance of inter-instance communication significantly, which could speedup distributed AI training and inference.

Using EFA on AWS VM#

For the instance types listed in the GPU and EFA count mapping table in the Adding EFA configurations in SkyPilot YAML section, the EFA can be enabled by setting resources.network_tier: best in the task YAML.

resources:
  network_tier: best

To run the NCCL test with EFA support with AWS VM:

sky launch -c efa efa_vm.yaml

Check the efa_vm.yaml for the complete SkyPilot cluster yaml configurations.

Benchmark results#

We compare the performance with and without EFA using NCCL test reports with the same resources (2x p4d.24xlarge, i.e. 2xA100:8).

The Speed-up 1 EFA vs 1 ENI column is calculated by busbw with 1 EFA Interface (GB/s) / busbw Non-EFA with 1 Network Interface (GB/s) and the Speed-up 4 EFA vs 1 ENI column is calculated by busbw with 4 EFA Interfaces (GB/s) / busbw Non-EFA with 1 Network Interface (GB/s).

Message Size

busbw with 4 EFA Interfaces (GB/s)

busbw with 1 EFA Interface (GB/s)

busbw Non-EFA with 1 Network Interface (GB/s)

Speed-up 1 EFA vs 1 ENI

Speed-up 4 EFA vs 1 ENI

8 B

0

0

0

-

-

16 B

0

0

0

-

-

32 B

0

0

0

-

-

64 B

0

0

0

-

-

128 B

0

0

0

-

-

256 B

0

0

0

-

-

512 B

0.01

0.01

0.01

1 x

1 x

1 KB

0.01

0.01

0.02

0.5 x

0.5 x

2 KB

0.02

0.02

0.03

0.6 x

0.6 x

4 KB

0.04

0.04

0.05

0.8 x

0.8 x

8 KB

0.09

0.08

0.08

1.0 x

1.1 x

16 KB

0.16

0.15

0.10

1.5 x

1.6 x

32 KB

0.28

0.26

0.16

1.6 x

1.7 x

64 KB

0.54

0.50

0.29

1.7 x

1.8 x

128 KB

1.07

0.81

0.45

1.8 x

2.4 x

256 KB

2.02

1.23

0.74

1.6 x

2.7 x

512 KB

3.28

1.70

0.85

2.0 x

3.8 x

1 MB

4.97

2.34

1.52

1.5 x

3.2 x

2 MB

6.77

2.80

2.35

1.2 x

2.9 x

4 MB

9.28

4.96

3.79

1.3 x

2.4 x

8 MB

10.99

8.21

5.37

1.5 x

2.0 x

16 MB

19.63

11.84

6.46

1.8 x

3.0 x

32 MB

32.62

14.93

7.27

2.0 x

4.5 x

64 MB

46.11

17.27

6.83

2.5 x

6.7 x

128 MB

57.79

18.67

7.93

2.3 x

7.3 x

256 MB

67.20

19.59

8.03

2.4 x

8.3 x

512 MB

72.90

19.99

8.14

2.4 x

8.9 x

1 GB

76.14

20.19

8.15

2.5 x

9.3 x

2 GB

77.90

20.31

8.20

2.5 x

9.5 x

From the above benchmark results, we can see that:

  • EFA brings little benefit for small messages, but gains grow with message size.

  • Bandwidth scales near-linearly with multiple EFAs, reaching ~78 GB/s with 4 interfaces.

  • On p4d.24xlarge (A100×8), 4 EFAs deliver up to ~9.5× higher bandwidth vs a single ENI.

So EFA is critical for scalable, high-throughput training workloads.

Included files#

efa_container.yaml

# This example is used to test the NCCL performance
# with EFA with container runtime on AWS VM.
name: nccl-efa-container

resources:
  infra: aws
  instance_type: p4d.24xlarge
  image_id: docker:public.ecr.aws/hpc-cloud/nccl-tests:latest
  network_tier: best

num_nodes: 2

run: |
  if [ "${SKYPILOT_NODE_RANK}" == "0" ]; then
    echo "Head node"

    # Total number of processes, NP should be the total number of GPUs in the cluster
    NP=$(($SKYPILOT_NUM_GPUS_PER_NODE * $SKYPILOT_NUM_NODES))

    # Append :${SKYPILOT_NUM_GPUS_PER_NODE} to each IP as slots
    nodes=""
    for ip in $SKYPILOT_NODE_IPS; do
      nodes="${nodes}${ip}:${SKYPILOT_NUM_GPUS_PER_NODE},"
    done
    nodes=${nodes::-1}
    echo "All nodes: ${nodes}"

    /opt/amazon/openmpi/bin/mpirun \
      --allow-run-as-root \
      --tag-output \
      -H $nodes \
      -np $NP \
      -N $SKYPILOT_NUM_GPUS_PER_NODE \
      --bind-to none \
      -x LD_LIBRARY_PATH \
      -x NCCL_DEBUG=INFO \
      --mca pml ^cm,ucx \
      --mca btl tcp,self \
      --mca btl_tcp_if_exclude lo,docker0,veth_def_agent \
      --mca plm_rsh_args "-p 10022" \
      /opt/nccl-tests/build/all_reduce_perf \
      -b 8 \
      -e 2G \
      -f 2 \
      -g 1 \
      -c 5 \
      -w 5 \
      -n 100
  else
    echo "Worker nodes"
  fi

efa_vm.yaml

# This example is used to test the NCCL performance
# with EFA on AWS VMs.
name: nccl-efa-vm

resources:
  infra: aws
  instance_type: p4d.24xlarge
  network_tier: best

num_nodes: 2

run: |
  if [ "${SKYPILOT_NODE_RANK}" == "0" ]; then
    echo "Head node"

    # Total number of processes, NP should be the total number of GPUs in the cluster
    NP=$(($SKYPILOT_NUM_GPUS_PER_NODE * $SKYPILOT_NUM_NODES))

    # Append :${SKYPILOT_NUM_GPUS_PER_NODE} to each IP as slots
    nodes=""
    for ip in $SKYPILOT_NODE_IPS; do
      nodes="${nodes}${ip}:${SKYPILOT_NUM_GPUS_PER_NODE},"
    done
    nodes=${nodes::-1}
    echo "All nodes: ${nodes}"

    /opt/amazon/openmpi/bin/mpirun \
      --allow-run-as-root \
      --tag-output \
      -H $nodes \
      -np $NP \
      -N $SKYPILOT_NUM_GPUS_PER_NODE \
      --bind-to none \
      -x LD_LIBRARY_PATH \
      -x NCCL_DEBUG=INFO \
      --mca pml ^cm,ucx \
      --mca btl tcp,self \
      --mca btl_tcp_if_exclude lo,docker0,veth_def_agent \
      /usr/local/cuda/efa/test-cuda-12.8/all_reduce_perf \
      -b 8 \
      -e 2G \
      -f 2 \
      -g 1 \
      -c 5 \
      -w 5 \
      -n 100
  else
    echo "Worker nodes"
  fi

nccl_efa.yaml

# This example is used to test the NCCL performance
# with EFA on HyperPod/EKS.
name: nccl-efa-eks

resources:
  infra: k8s
  accelerators: A100:8
  cpus: 90+
  image_id: docker:public.ecr.aws/hpc-cloud/nccl-tests:latest

num_nodes: 2

envs:
  USE_EFA: "true"

run: |
  if [ "${SKYPILOT_NODE_RANK}" == "0" ]; then
    echo "Head node"

    # Total number of processes, NP should be the total number of GPUs in the cluster
    NP=$(($SKYPILOT_NUM_GPUS_PER_NODE * $SKYPILOT_NUM_NODES))

    # Append :${SKYPILOT_NUM_GPUS_PER_NODE} to each IP as slots
    nodes=""
    for ip in $SKYPILOT_NODE_IPS; do
      nodes="${nodes}${ip}:${SKYPILOT_NUM_GPUS_PER_NODE},"
    done
    nodes=${nodes::-1}
    echo "All nodes: ${nodes}"

    # Set environment variables
    export PATH=$PATH:/usr/local/cuda-12.2/bin:/opt/amazon/efa/bin:/usr/bin
    export LD_LIBRARY_PATH=/usr/local/cuda-12.2/lib64:/opt/amazon/openmpi/lib:/opt/nccl/build/lib:/opt/amazon/efa/lib:/opt/aws-ofi-nccl/install/lib:/usr/local/nvidia/lib:$LD_LIBRARY_PATH
    export NCCL_HOME=/opt/nccl
    export CUDA_HOME=/usr/local/cuda-12.2
    export NCCL_DEBUG=INFO
    export NCCL_BUFFSIZE=8388608
    export NCCL_P2P_NET_CHUNKSIZE=524288
    export NCCL_TUNER_PLUGIN=/opt/aws-ofi-nccl/install/lib/libnccl-ofi-tuner.so

    if [ "${USE_EFA}" == "true" ]; then
      export FI_PROVIDER="efa"
    else
      export FI_PROVIDER=""
    fi

    /opt/amazon/openmpi/bin/mpirun \
      --allow-run-as-root \
      --tag-output \
      -H $nodes \
      -np $NP \
      -N $SKYPILOT_NUM_GPUS_PER_NODE \
      --bind-to none \
      -x FI_PROVIDER \
      -x PATH \
      -x LD_LIBRARY_PATH \
      -x NCCL_DEBUG=INFO \
      -x NCCL_BUFFSIZE \
      -x NCCL_P2P_NET_CHUNKSIZE \
      -x NCCL_TUNER_PLUGIN \
      --mca pml ^cm,ucx \
      --mca btl tcp,self \
      --mca btl_tcp_if_exclude lo,docker0,veth_def_agent \
      /opt/nccl-tests/build/all_reduce_perf \
      -b 8 \
      -e 2G \
      -f 2 \
      -g 1 \
      -c 5 \
      -w 5 \
      -n 100
  else
    echo "Worker nodes"
  fi

config:
  kubernetes:
    pod_config:
      spec:
        containers:
        - resources:
            limits:
              vpc.amazonaws.com/efa: 4
            requests:
              vpc.amazonaws.com/efa: 4