Cases7 cases

1 to 5 of 7

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-0129 Throughput of the 384-GPU SmolLM3 pretraining run collapsed because shared WekaFS storage evicted dataset shards to S3, then kept dropping on local NVMe because the nanotron dataloader built an index that grew with the step count

   Hugging Face2025

   Symptoms

   Within hours of launching the full SmolLM3 pretraining run (3B parameters, 11T tokens, 384 H100 GPUs on 48 nodes), training throughput fell steeply, with repeated sharp drops. The team expected a steady 13,500 to 14,000 tokens per second, and none of the throughput ablations run beforehand on the same 48 nodes had shown drops.

   Throughput degradation

   Trigger

   Launching the full run with the whole ~24 TB dataset on shared WekaFS storage and the real step count for 11T tokens, both far larger than in the ablation runs.

   Root cause

   Capacity exhaustion

   Component

   Shared network filesystemTraining data loader

   Contributing factor

   Overlapping failuresUnrepresentative stagingUnverified assumption

   Detection

   The team watched training throughput in tokens per second and saw it fall within hours of launch; node-monitoring metrics then tied the large fall to spikes in disk read latency.

   Remediation

   • moved the training data from WekaFS to the local NVMe RAID scratch storage on each node
   • copied the dataset to replacement nodes from a healthy node with fpsync, cutting the copy from 3 hours to 1.5 hours
   • kept a spare node with the dataset preloaded in the Slurm reservation to swap in when a node failed
   • reproduced the drops on a single node and with step counts of about 3M and 32k to rule out hardware and isolate the step count
   • replaced nanotron's nanosets dataloader with the TokenizedBytes dataloader copied from the SmolLM2 setup, which removed the drops
   • pre-shuffled the tokenized sequences offline for each epoch with different seeds (about 1 hour) to fix the noisier loss from TokenizedBytes

   Lessons

   Ablations on data subsets and a short step horizon did not exercise either cause: a shared cache-and-evict filesystem near its capacity, and a dataloader index that grows with the step count. Reproducing the drops on one node separated hardware from software, and varying only the step count pointed at the dataloader. The team chose proven replacements over repairs under deadline pressure, taking an existing dataloader instead of fixing nanosets and offline shuffling instead of changing the dataloader, and warns that too many such shortcuts leave a patchwork system that is hard to maintain.

   View case
3. 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
4. IL-0128 Network interfaces that dropped for a few seconds made NCCL time out and stalled or slowed LLM training on a cluster of more than 10,000 GPUs

   ByteDance2024

   Symptoms

   ByteDance's large language model training jobs on its MegaScale system, which runs on more than 10,000 GPUs, occasionally stalled or slowed down. The cause was network interfaces that flapped frequently.

   Link flappingNCCL timeoutThroughput degradation

   Trigger

   Network interfaces went down for several seconds and came back up, dropping every packet in transit.

   Root cause

   DNS or network fault

   Component

   NCCLNIC, cable and switch port link

   Contributing factor

   Unreviewed defaults

   Detection

   MegaScale's monitoring system uses second-level monitoring to track link flapping and other NIC issues. When the RDMA traffic metrics carried in the training daemons' heartbeats drop sharply or fluctuate abnormally, the driver raises an alert for manual investigation, and if the traffic stops entirely it starts the fault recovery procedure automatically.

   Remediation

   • set NCCL's timeout threshold explicitly to a larger value than the default
   • tuned NCCL's retransmit timer and retry count for fast recovery under link flapping
   • enabled the adap_retrans feature on the NICs
   • applied lower-level quality control to NIC signal strength, AOC cable quality and switch-side signal strength

   Lessons

   A default NCCL timeout can be shorter than the few seconds an interface needs to come back from a flap, so a brief link loss becomes a failed collective and a stalled job. A larger timeout and faster retransmission give the link time to return, but the flapping itself came from physical link quality between NIC, AOC cable and switch, and quality control at that level reduced it.

   View case
5. IL-0198 An automatic Ubuntu security update restarted systemd-networkd, which flushed the IP rules Cilium relied on and cut more than 60 percent of Datadog's Kubernetes nodes off the network across regions and cloud providers

   Datadog2023

   Symptoms

   Between 06:00 and 07:00 UTC on 8 March 2023, Kubernetes nodes in several Datadog regions lost network connectivity at the same time, on AWS, Google Cloud and Azure. All services across multiple regions were affected, including the regions that host Datadog's CI/CD and automation tools. Datadog's regions are not connected to each other and changes are rolled out one region at a time, so a simultaneous failure on several cloud providers looked implausible at first.

   Node NotReadyService outageThroughput degradation

   Trigger

   On 7 March 2023 Ubuntu published a systemd patch for a CVE (249.11-0ubuntu3.7). Ubuntu's default unattended upgrades installed it on the nodes between 06:00 and 07:00 UTC on 8 March, and installing it restarted systemd-networkd.

   Root cause

   Breaking change in an upgrade

   Component

   systemd-networkdCiliumNode operating system image

   Contributing factor

   Unreviewed defaultsChange with a delayed effectNo staged rolloutAutomation amplification

   Remediation

   • restarted thousands of disconnected instances on Google Cloud and Azure through the cloud APIs, keeping their local disks
   • rebuilt the data lost with terminated AWS instances' local disks through replication, reprocessing and restores from backup

   Lessons

   A change in a base operating system component can stay dormant for months until something restarts that component. Here the trigger was a security patch, and the change reached every region at once because automatic OS updates ran outside the progressive, region-by-region process used for every other change. Cloud providers react differently to an instance that loses its network. Google Cloud and Azure left the instances running, so a restart recovered them with their disks; AWS replaced them quickly, which brought stateless services back fast but destroyed the local-disk data the stateful services needed.

   View case

By root cause6