HyperPod Node Debugger
Operating policy. Run read-only diagnostics yourself. Never run a command that changes cluster, node, or workload state — present each one as a Suggested command (run this yourself) block and wait for the customer. Destructive order: investigate → reboot → replace (replace destroys root + secondary volumes; not supported on Slurm controller nodes). Never discard training state, logs, or caches on speculation.
IaC note (always include with mutation commands). When you suggest any command that changes cluster, VPC, SG, subnet, or EKS configuration (e.g. authorize-security-group-*, modify-vpc-attribute, update-cluster, kubectl label/cordon/drain, create namespace, set env daemonset), ask the customer first whether the cluster / VPC / SG is managed by Infrastructure-as-Code (CloudFormation, CDK, Terraform, Pulumi). If yes, tell them: "Apply this change in your IaC source first, then deploy through the pipeline — running the command directly will drift from your template and the next stack update may overwrite it." If they need to fix the issue immediately and the IaC change will follow, flag the drift explicitly so they remember to reconcile.
Read-only triage. scripts/triage-cluster.sh (and helpers check-efa-sg.sh, check-node-reachability.sh, check-vpc-config.sh) read state and print each issue as [FAIL] ... → references/node-diagnostics-detail.md § <section>. Catalog of customer-ticket patterns: references/node-issue-catalog.md.
Workflow
- Collect cluster name, region, suspect instance ID, exact error string from logs.
- Run
scripts/triage-cluster.sh (add --node <INSTANCE-ID> to focus one node).
- For every
[FAIL] / issue entry, Read the referenced section.
- Present: what script detected (copy the line verbatim), root cause, exact command(s) with instance/SG IDs filled in, blast radius (e.g. "reboots i-xxx", "wipes volumes on replacement"). For any command that mutates cluster/VPC/SG/EKS state, ask whether the affected resource is IaC-managed and surface the drift warning from the operating-policy note above.
- Wait for explicit customer approval. Destructive order: investigate → reboot → replace.
- Re-run triage to confirm. Iterate if not cleared.
Step 1: Triage
bash scripts/triage-cluster.sh --cluster <CLUSTER_NAME_OR_ARN> --region <REGION>
# Focus on one node:
bash scripts/triage-cluster.sh --cluster <CLUSTER_NAME_OR_ARN> --region <REGION> --node <INSTANCE_ID>
One pass collects: cluster status + NodeRecovery, events, per-node health (HyperPod + EKS labels, Slurm states), VPC/SG snapshot, CloudWatch availability, SSM readiness, on-node resource checks (disk, memory, /dev/shm, OOM, NVMe, time sync, SSM agent), Slurm node→instance mapping.
Tags: [PASS] passed · [FAIL] issue with a → references/... pointer · [WARN] advisory · [INFO] informational. Priorities: P0 blocks operation · P1 degraded · P2 informational.
Step 2: Match signal → section
Events (list-cluster-events) — provisioning-time:
| Event | Section |
|---|
"EFA health checks did not run successfully" (public-doc verbatim signal) | A: EFA/SG |
| Instance bootstrap or network-misconfiguration event | A + B: VPC |
| Lifecycle-script failure or timeout | D: Lifecycle |
| Insufficient-capacity or AZ-mismatch failure at creation | C: Capacity |
Hardware failure / UnschedulablePendingReplacement | F: Hardware |
EKS labels:
| Label | Section |
|---|
node-health-status: UnschedulablePendingReplacement | F |
node-health-status: UnschedulablePendingReboot | F |
deep-health-check-status: Failed | G → F |
Symptoms:
| Symptom | Section |
|---|
| Training hangs at NCCL init / AllReduce | A → E |
Slurm node down / "Node unexpectedly rebooted" | H: Slurm |
| Jobs stuck PENDING / COMPLETING | H |
| Auto-repair not triggering | F |
| GPU not visible / XID / ECC errors | G |
| GPU row-remap pending/failed / silent NaNs / DCGM Fail | G § G.1.a/b |
Disk full / OOM / "Cannot allocate memory" | I: Resources |
| Wrong vCPU count (e.g. 96 instead of 192 on p5.48xlarge) | J: Config |
| Container CrashLoopBackOff / runtime crash | M: Container Runtime |
aws-node CrashLoopBackOff / gRPC 50051 refused | O: CNI / Pod Networking |
| Pods stuck Pending with no IP / CNI error | O |
DNS resolution / enableDnsSupport | B § B.2 |
| Public subnet / IGW misconfigured | B § B.3 |
| Missing VPC endpoints (ECR / STS / FSx) | B § B.4 |
| EKS VPC / SG mismatch with HyperPod | B § B.5 |
| Kernel panic / watchdog / hung task | N: Kernel |
| Need shell on a node | K: SSM |
| Collect logs for AWS Support | L: Log Collection |
A: EFA / Security Group
Per the HyperPod prerequisites doc, the SG must allow all inbound and outbound to itself. scripts/check-efa-sg.sh validates self-ref rules on every cluster SG. On-node EFA check via scripts/check-node-reachability.sh over SSM. Full: § A.
B: VPC / Routing
SG/subnet VPC mismatch, missing S3 Gateway endpoint, EKS auth mode, worker→controller routing, VPC DNS support, private-subnet + NAT / VPC endpoints, EKS↔HyperPod VPC alignment. scripts/check-vpc-config.sh. Full: § B.
C: Capacity / AZ
Insufficient-capacity failure at creation, or no subnets in the AZ where capacity is available. Check AZ offerings via describe-instance-type-offerings, then change subnet AZ or use Flexible Training Plans / ODCR. Full: § C.
D: Lifecycle Scripts
Surfaced in cluster events + CloudWatch under LifecycleConfig/<group>/<instance-id>. Common: S3 connectivity, IAM gaps, CRLF line endings, infinite loops, parameter-name mismatch. Full: § D.
E: Software Versions
Delegate to hyperpod-version-checker to compare NVIDIA driver, CUDA, NCCL, EFA installer, OFI NCCL, PyTorch across nodes. Ensure job env has FI_PROVIDER=efa, FI_EFA_USE_DEVICE_RDMA=1, NCCL_SOCKET_IFNAME=^lo,docker. Full: § E.
F: Hardware / Auto-Repair
Confirm NodeRecovery=Automatic, inspect the EKS health labels + sagemaker.amazonaws.com/fault-details annotation, and read the SagemakerHealthMonitoringAgent/<group>/<instance> CloudWatch stream. HMA runs passive background checks on GPU and Neuron state and reboots the node on count mismatch (per the HMA doc: "if there's a mismatch between the expected number of GPUs … and the count returned by nvidia-smi, then HMA reboots the node"; same for neuron-ls). Manual recovery order: reboot first, replace only if reboot fails; the preferred path is the batch APIs (BatchReboot/BatchReplaceClusterNodes). Full: § F · patterns: node-issue-catalog.md.
G: GPU / Accelerator
NVIDIA (p4d/p5/g5/g6): nvidia-smi + dmesg over SSM for Xid, ECC, thermal throttling. Xid classification per NVIDIA's catalog: 13 Graphics Engine Exception (application-level), 31 GPU memory page fault (application, can be driver/HW), 63 GPU memory remapping event (HW/ECC), 71 CE4 Error (HW copy engine), 74 NVLink Error (HW), 79 GPU has fallen off the bus (PCIe bus), 109 Context Switch Timeout Error (HW). Any uncorrectable ECC → drain and replace. Row-remap state is the authoritative silent-degradation signal (§ G.1.a).
Trainium / Inferentia (trn1/trn2/inf2): Neuron SDK — neuron-ls, neuron-top, neuron-monitor. nvidia-smi does not apply.
GPU / accelerator failures flow into § F for reboot / replace. Full: § G.
H: Slurm Node Management
Node down/unresponsive, unexpected reboots, stuck PENDING/COMPLETING jobs, Slurm-to-instance-ID translation. Primary access is SSM; diagnose slurmd first, fix the root cause, then start/resume the node per § H. Full: § H.
I: Resource Exhaustion
Disk full (HyperPod root volume defaults to 100 GB and is not intended to grow post-creation), OOM, os.fork() memory error, /dev/shm exhaustion, inode exhaustion. Fork-memory fix: export FI_EFA_USE_HUGE_PAGE=0. Redirect bulk data to /opt/sagemaker (secondary EBS) or /opt/dlami/nvme (instance store). Full: § I.
J: Configuration
p5.48xlarge reports 96 vCPU instead of 192 → set ThreadsPerCore=2 via update-cluster. Full: § J.
K: Node Access via SSM
No direct SSH on HyperPod. Target format sagemaker-cluster:<CLUSTER_ID>_<GROUP>-<INSTANCE_ID>. Failures: plugin missing, wrong prefix, IAM, VPC endpoints. Full: § K.
L: Log Collection
Delegate to hyperpod-issue-report for S3-stored bundles. Key CloudWatch streams: LifecycleConfig/<group>/<instance-id>, SagemakerHealthMonitoringAgent/<group>/<instance-id>. Full: § L.
M: Container Runtime
CrashLoopBackOff, OOMKilled, ImagePullBackOff, RunContainerError on EKS. kubectl describe pod + on-node crictl ps -a, journalctl -u containerd. Full: § M.
N: Kernel & System
Kernel panic, watchdog timeout, soft lockup, unexpected reboots not explained by HyperPod health monitoring. dmesg | grep -iE 'panic|watchdog|hung_task|NMI' + journalctl -b -1. nvrm-related signatures point at NVIDIA driver crashes. Full: § N.
O: CNI / Pod Networking
VPC CNI (aws-node) failures, IPAMD errors, gRPC 127.0.0.1:50051 refused, pods stuck Pending with FailedCreatePodSandBox. Script auto-checks aws-node, kube-proxy, CoreDNS. Full: § O.
Prerequisites
aws CLI v2, recent enough to support the HyperPod cluster commands (describe-cluster, list-cluster-nodes, batch-reboot-cluster-nodes, batch-replace-cluster-nodes)python3, bash 4+ (associative arrays are required by the scripts)kubectl authenticated to the EKS cluster (K8s checks skipped if absent)session-manager-plugin for on-node hardware checksunbuffer (from the expect package) — optional; if missing, SSM on-node probes are skipped while the rest of the triage still runs. Install via yum install expect / apt install expect.
Defaults
- Region — required: pass
--region or set $AWS_DEFAULT_REGION.
- Target scope — all nodes;
--node <ID> focuses one.
- Event window — up to 500 most recent events (5 × 100, paginated).
- Node list cap — up to 20,000 nodes (200 × 100); warns on cap.
- SSM probes — 180 s per node with retry-on-throttle.
- Colors — auto-disabled on non-TTY;
--no-color to force off.
Error handling
| Failure | Script | Tell the customer |
|---|
aws sts get-caller-identity fails | Exit 1 | "Fix AWS credentials and rerun." |
describe-cluster fails | Exit 1 after listing region's clusters | "Confirm cluster name and region." |
sagemaker:* / ec2:* / logs:* AccessDenied | Warn, add Missing IAM permission for <API>, continue | "Grant the listed IAM action and rerun." |
kubectl absent or unauthenticated | Skip K8s checks | "Install/authenticate kubectl (see § K)." |
session-manager-plugin absent | Skip on-node probes | "Install session-manager-plugin (see § K)." |
SSM start-session fails or times out (180s) | Mark node unreachable with → § K pointer | "Rerun with --node <ID> to isolate; verify SSM agent on the node." |
| Cluster > 20,000 nodes | First 20,000 paginated; warn | "Use --node to target specific nodes." |
Exit codes: 0 triage complete · 1 cluster not found or fatal prerequisite missing.
IAM permissions
Read-only diagnostic — covers triage-cluster.sh, check-efa-sg.sh, check-vpc-config.sh, and check-node-reachability.sh:
{
"Action": [
"sagemaker:DescribeCluster",
"sagemaker:DescribeClusterNode",
"sagemaker:ListClusterNodes",
"sagemaker:ListClusterEvents",
"sagemaker:ListClusters",
"eks:DescribeCluster",
"ec2:DescribeSecurityGroups",
"ec2:DescribeSubnets",
"ec2:DescribeVpcs",
"ec2:DescribeVpcAttribute",
"ec2:DescribeVpcEndpoints",
"ec2:DescribeRouteTables",
"ec2:DescribeNetworkInterfaces",
"ec2:DescribeInstances",
"ec2:DescribeInstanceTypeOfferings",
"ec2:DescribeInstanceTypes",
"logs:DescribeLogGroups",
"logs:DescribeLogStreams",
"logs:FilterLogEvents",
"ssm:StartSession",
"ssm:TerminateSession",
"service-quotas:GetServiceQuota"
]
}
sts:GetCallerIdentity is implicit — it requires no IAM action. SSM on HyperPod uses start-session against sagemaker-cluster:<cluster-id>_<group>-<iid> targets — not send-command against bare instance IDs. For remediation commands, grant the matching write permission (e.g. ec2:AuthorizeSecurityGroupIngress / Egress, ec2:RevokeSecurityGroupIngress / Egress, ec2:ModifyVpcAttribute, sagemaker:UpdateCluster, sagemaker:BatchRebootClusterNodes, sagemaker:BatchReplaceClusterNodes). Not needed for the diagnostic itself.
Skill delegation
| Need | Use |
|---|
| Cluster creation / deployment failures | hyperpod-cluster-debugger (§ A / B / C / H + --validate) |
| Cluster-wide SSM outage | hyperpod-cluster-debugger § F |
| Single-node SSM failure | stay here — § K |
| Cluster-wide EFA health-check failure at creation time | hyperpod-cluster-debugger § A |
| Single-node EFA failure post-provisioning | stay here — § A |
| NCCL AllReduce / collective-op timeouts (distributed) | hyperpod-nccl |
| Silent GPU NaNs on a specific node (row-remap / DCGM) | stay here — § G.1 (even if discovered by NCCL) |
| Post-deployment cluster-wide management | hyperpod-cluster-debugger |
| Shell / commands on nodes | hyperpod-ssm |
| CUDA / NCCL / EFA version comparison | hyperpod-version-checker |
| Diagnostic bundle for AWS Support | hyperpod-issue-report |
| Training performance / MFU degradation | hyperpod-mfu-debugger |
Escalate to AWS Support
Escalate when:
- SG rules correct and reachability passes but EFA still fails.
- VPC correct but K8s bootstrap fails — check VPC flow logs for REJECT.
- Hardware failure where replacement keeps failing (bad physical host).
- Node replacement fails with an insufficient-capacity signal despite a valid ODCR.
Before opening the case
# 1. Cluster identity + affected node status
aws sagemaker describe-cluster --cluster-name <CLUSTER> --region <REGION>
aws sagemaker list-cluster-nodes --cluster-name <CLUSTER> --region <REGION> \
--query "ClusterNodeSummaries[?InstanceId=='<INSTANCE_ID>']"
# 2. Triage bundle (scoped to the affected node where possible)
bash scripts/triage-cluster.sh --cluster <CLUSTER> --region <REGION> --node <INSTANCE_ID> > triage.txt
# 3. Per-node log/config bundle to S3 (delegates to hyperpod-issue-report)
# See skills/hyperpod-issue-report/SKILL.md for the exact invocation.
Include in the case
- Cluster name + ARN and AWS region
- Orchestrator (EKS or Slurm)
- Affected instance IDs / node names / instance-group names
- Timestamp window (UTC start / end) of the failure
- Exact error strings observed (copy verbatim from pod logs, CloudWatch, dmesg, events)
- XID numbers / ECC counts / DCGM output where hardware is implicated
triage.txt from step 2 above- S3 URI of the
hyperpod-issue-report bundle from step 3
Patterns from real customer tickets: node-issue-catalog.md.