Diagnose and configure AWS EFA for distributed GPU training on EKS.
npx skills add https://github.com/tylertitsworth/skills --skill aws-efa
Please help me install this Agent Skill. Skill: aws-efa Source: https://github.com/tylertitsworth/skills/tree/main/aws-efa Command: npx skills add https://github.com/tylertitsworth/skills --skill aws-efa
AWS Elastic Fabric Adapter (EFA) provides a high-bandwidth, low-latency network path for distributed GPU workloads by enabling SRD/libfabric-based OS-bypass and GPUDirect RDMA, reducing CPU overhead and enhancing cross-node communication on EKS.
Create an EFA-enabled EKS nodegroup and verify NCCL throughput using the aws-ofi-nccl provider.
High-intent search queries and answers about installing and using this skill.
To enable EFA for multi-node GPU training on EKS, you must create an EFA-enabled nodegroup and configure the aws-ofi-nccl provider. This establishes SRD and libfabric-based OS-bypass communication, reducing CPU overhead for cross-node workloads.
AWS Elastic Fabric Adapter provides a high-bandwidth, low-latency network path for distributed GPU workloads by enabling SRD and libfabric-based OS-bypass. It utilizes GPUDirect RDMA to enhance cross-node communication on EKS clusters.
Tuning NCCL over EFA requires understanding ENA and EFA topology alongside aws-ofi-nccl integration. You must configure node settings specifically for GPUDirect RDMA workloads to optimize multi-node NCCL communication on AWS GPU clusters.
Cross-node NCCL communication failures on AWS GPU clusters often stem from incorrect EFA topology, security group configurations, or placement group settings. Troubleshooting requires verifying your aws-ofi-nccl integration and SRD protocol setup.
Yes, EFA-enabled NCCL deployments on EKS require specific placement groups to ensure optimal network topology. Proper placement groups guarantee low-latency SRD protocol communication and effective GPUDirect RDMA across cross-node GPU clusters.
sitharaj88Generate AWS EKS cluster configurations, node groups, Helm charts, and Kubernetes manifests.
tomzConfigure CoreWeave GPU clusters with node selectors, MIG profiles, and NCCL tuning.
haozhx23Checks and compares software component versions on SageMaker HyperPod cluster nodes.
wcyganDiagnoses Talos OS health and node networking via read-only queries across all nodes.
