Cases4 cases

1 to 4 of 4

1. IL-0137 Training initialisation on a 60-node NVIDIA B200 cluster took more than 8 hours instead of minutes because fragmented small random I/O from all nodes saturated the metadata service of the VAST Data NFS storage

   Lablup Inc2026

   Symptoms

   The 60-node NVIDIA B200 training configuration for the Solar Open pre-training run showed no major problem during initial validation on 2 to 4 nodes. Once the job ran at its full size of 60 nodes, training initialisation, which reads from the VAST Data storage and should have finished within minutes, took more than 8 hours, and storage throughput stayed far below the theoretical limit.

   Throughput degradation

   Trigger

   Scaling the training job from 2 to 4 nodes in validation to the full 60-node configuration.

   Root cause

   Capacity exhaustion

   Component

   Shared network filesystem

   Contributing factor

   Unrepresentative stagingMonitoring blind spot

   Detection

   The problem showed up as training initialisation taking hours after the job was scaled to 60 nodes; application logs showed only the delayed initialisation.

   Remediation

   • Sharded the Arrow training data files by rank on the application side
   • Switched the storage to asynchronous deletion instead of synchronous rm/unlink
   • Tuned readahead

   Lessons

   Storage behaviour measured on 2 to 4 nodes did not predict behaviour on 60 nodes: a metadata bottleneck of this kind appears only at production scale, so small pre-tests cannot find it. Per-node network rates can look normal while the combined access pattern of all nodes overwhelms the storage metadata path. Finding the cause took metrics from the application, scheduler, network and storage on one shared timeline; monitoring by any single team was not enough. The report concludes that GPU capacity has to be provisioned together with a matching storage path.

   View case
2. IL-0025 Failed boots, InfiniBand link flapping, NCCL hangs and throughput sags while bringing up a 4,088-GPU H100 training cluster

   Imbue2024

   Symptoms

   When Imbue brought up a cluster of 4,088 H100 GPUs on 511 hosts, about 10% of the machines failed to boot, mostly from physical problems: unconnected or miswired Ethernet cables, iDRAC hardware issues, broken power supplies, bad NVMe drives, missing internal wires, and network cards or GPUs that did not show up. In single-node GPU tests, Ubuntu server logs showed links between the GPUs and the PCIe buses or network cards reporting "limited width: x4 < x16". A few hosts had NVMe drives that locked up the whole machine when touched, drives that appeared in random order and got the OS installed on the wrong one, wrong temperature readings that kept fans at 100%, and CPUs limited to 2 GHz.

   Link flappingNCCL timeoutThroughput degradationGPU unavailableTraining job failure

   Trigger

   Bringing a new InfiniBand-connected H100 cluster into service and running the first multi-node training jobs on it.

   Root cause

   Dependency failure

   Component

   InfiniBand fabricNCCLGPU driverPyTorch

   Contributing factor

   Unverified assumptionMonitoring blind spotPartial rollout

   Detection

   Automated hardware checks during provisioning, UFM alerts on temperature, error rates and flapping ports, Prometheus node exporter and DCGM metrics, per-batch MFU monitoring, and live debugging of stalled processes with Py-Spy, GDB and NVIDIA Nsight.

   Remediation

   • Automated checks for broken machines and sent faulty ones back to Dell or to data center staff
   • Reseated GPUs and internal PCIe cables and applied firmware upgrades
   • Downgraded the NVIDIA driver that caused wrong temperature readings
   • Rewired the top tier of the InfiniBand fabric into a single network and added checks that the wiring matched the design
   • Cleaned and reseated alerting ports and disabled the remaining alerting transceivers
   • Ran a burn-in workload through every InfiniBand port and disabled the ports that did not survive it
   • Enabled GPUDirect RDMA with an extra kernel module and PCIe ACS disabled
   • Disabled automatic garbage collection and scheduled it at set intervals on all hosts

   Action items

   • Release host-level health checks that confirm a host is free from known errors
   • Release an NCCL patch that improves logging around errors and stalls
   • Release a GPU stress test and NVLink and InfiniBand networking tests
   • Release a script that parses the UFM event log and decides which network ports to disable
   • Release a script that generates a burn-in workload exercising every InfiniBand link
   • Relaunch crashed runs automatically from the last checkpoint on the machines that pass health checks
   • Replace the disabled alerting transceivers

   Lessons

   Keep 10 to 20% more machines than a run needs, and network them so any working subset can be used. On the newest hardware expect about 3% of machines to break each week, mostly the same few machines breaking in different ways, so growing a set of known reliable "golden" hosts works better than chasing random failures. Write tests and automation for every kind of failure met, because each one comes back, and write tools that make opaque errors readable. Change only one thing at a time, and double-check the claims of new tools and people before building on them.

   View case
3. IL-0116 A CNI configuration change incompatible with a hardware variant that exists only in production cut network connectivity in an OpenAI GPU cluster and took text-davinci-003 fully down

   OpenAI2023

   Symptoms

   At 4:53 pm Pacific on 25 January 2023 one of OpenAI's GPU clusters had a major outage, and the text-davinci-003 model went fully down.

   Service outageThroughput degradation

   Trigger

   A configuration change to the CNI (Container Network Interface) plugins, which provide network connectivity to containers, was deployed to a GPU cluster that contained a hardware variant the staging environment did not have.

   Root cause

   Configuration change

   Component

   CNI plugin

   Contributing factor

   Unrepresentative stagingMonitoring blind spot

   Detection

   Engineers immediately identified the problem as network connectivity, but it took nearly an hour to identify CNI as the cause.

   Remediation

   • put a fix for the CNI configuration change in place, which restored 80% of traffic immediately
   • deployed models back onto GPUs over about three more hours until capacity was fully stable

   Action items

   • Make staging environments contain every type of hardware configuration so they better represent production
   • Monitor and alert on the health of the CNI system directly
   • Fix the network and storage bottlenecks that limit how quickly many replicas of a model can be deployed

   Lessons

   A change that had been deployed to other clusters over 24 hours and deemed safe still broke a cluster with different hardware, so staging needs every hardware configuration that production has.

   View case
4. IL-0032 Mutating admission controller generated broken JSON patches, silently blocking deploys for a week until a production service ran down to two pods

   Airbnb2020

   Symptoms

   A production service at Airbnb showed a high error rate and could not be deployed. The deploy failed with a cryptic Kubernetes error about being unable to access invalid index 5.

   High error rateThroughput degradationDeploys failing

   Trigger

   A change to the mutating admission controller, which had been rolled out within the previous month, had gone in seven days before the incident surfaced; when it had to delete two items from an array in a spec, it generated an invalid JSON patch that removed index 2 and then index 3.

   Root cause

   Software defect

   Component

   Admission webhookHorizontal pod autoscaler

   Contributing factor

   Change with a delayed effectMonitoring blind spotIgnored early signal

   Detection

   The affected service reported a high error rate and its deploy failed; the slow degradation of other services over the previous seven days had gone unnoticed.

   Remediation

   • deleted the mutating admission controller from the affected cluster

   Action items

   • Create alerts from the existing metrics that would have shown the problem earlier

   Lessons

   Keep in mind that a mutating admission controller exists in the cluster; a recently added one is easy to forget during an incident. Search for the exact error message early instead of working on the visible symptoms. Page a teammate when progress stalls.

   View case

By root cause4