NIXL Firmware Page Registration Fan-Out Triggers Host OOM Kills on HGX H200 and B200

NVIDIA Inference Xfer Library workloads on HGX H200 and HGX B200 instances triggered host OOM kills because UCX registered GPU memory as firmware pages across all 8 NICs simultaneously, consuming ~34 GiB of host memory per GPU in firmware-level registrations. The root cause was traced to UCX registering every reachable HCA for each buffer, creating firmware page fan-out proportional to NIC count.

NIXL Firmware Page Registration Fan-Out Triggers Host OOM Kills on HGX H200 and B200 is a Reliability failure seen during ML training runs. NVIDIA Inference Xfer Library workloads on HGX H200 and HGX B200 instances triggered host OOM kills because UCX registered GPU memory as firmware pages across all 8 NICs simultaneously, consuming ~34 GiB of host memory per GPU in firmware-level registrations. The root cause was traced to UCX registering every reachable HCA for each buffer, creating firmware page fan-out proportional to NIC count. Common tags: Crusoe, Nixl, Ucx, Firmware Page.

UCX registers each GPU buffer on every reachable HCA (all 8 NICs per node) creating firmware page duplication factor of N_nics. Each firmware page consumes ~35 KB of locked host memory per registration entry. With 75 GB registered per GPU * 8 NICs, firmware overhead reaches ~34 GiB of host memory before any workload memory. Taken together, these mechanisms explain why the failure is reproducible, why it tends to surface on specific workloads or scales, and why generic mitigation attempts often fall short without addressing the underlying cause.

• Host processes killed by OOM killer with GPUs at normal HBM utilization levels
• MemAvailable in /proc/meminfo drops to zero despite adequate host RAM capacity
• Swap fills up and kernel OOM killer activates without GPU memory pressure

• NVIDIA HGX H200 and HGX B200 with 8 NICs per node
• NIXL workloads with GPU memory registration exceeding ~75 GB per GPU
• UCX-based communication (UCX 1.20 and earlier)
• KVM-virtualized GPU instances with PCIe passthrough

Use this checklist to test the hypothesis against a small reproduction. No single line proves the root cause, so preserve the preceding events and compare one variable at a time.

• ✓Reproduce the failure from a clean checkpoint/seed: the symptom must appear without warm-up state from a previous run.
• ✓Verified signal present: /proc/meminfo shows MemAvailable = 0 with buffers/cache fully reclaimed during NIXL-heavy workloads
• ✓Verified signal present: fw_pages_total in debug counters shows >1 million firmware pages per NIC for ~75 GB registered per GPU
• ✓Verified signal present: Kernel dmesg: "Out of memory: Killed process <PID> (python)" with no GPU-side errors
• ✓A targeted fix from the "How to fix it" section eliminates or substantially reduces the symptom within one validation pass.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

• UCX registers each GPU buffer on every reachable HCA (all 8 NICs per node) creating firmware page duplication factor of N_nics
• Each firmware page consumes ~35 KB of locked host memory per registration entry
• With 75 GB registered per GPU * 8 NICs, firmware overhead reaches ~34 GiB of host memory before any workload memory