You need to enable JavaScript to run this app.
Volcengine Kubernetes Engine

Volcengine Kubernetes Engine

Copy page
Download PDF
Heterogeneous computing FAQs
How do I deal with the error in an RDMA-enabled environment?
Copy page
Download PDF
How do I deal with the error in an RDMA-enabled environment?
Issue phenomenon​
An NCCL runtime error occurs when containers use RDMA devices to achieve communication among multiple machines.​
  • ibv_create_cq allocate memory failed.​
  • ibvwrap.c:106 NCCL WARN Call to ibv_reg_mr failed.​
  • ​
    ibvwrap.c:106 NCCL WARN Call to ibv_reg_mr failed​
    proxy.cc:1519 NCCL WARN [Proxy Service 5] Failed to execute operation Connect from rank 5, retcode 2​
    NCCL WARN socketProgress: Connection closed by remote peer master-0.svc.cluster.local<36167>​
    ​
Environment configuration​
  • Node model: ecs.hpcpni3ln.45xlarge​
  • NCCL version: 2.18.5​
Cause analysis​
It might be caused by "the default size set by the command ulimit -l is 64K, which is insufficient for memlock memory in benchmark scenarios". In that case, please refer to the following solution.​
Solution​
​
Fix it on new nodes
Fix it on existing nodes
​
Go to the page "Edit node pool" and select "Advanced settings", then copy the following script to the textbox "InitializeScript" to overwrite all the existing content. Save the changes to take effect.​
Warning​
The value of the parameter LimitNOFILE in the script needs to be consistent with that of the same parameter in /lib/systemd/system/containerd.service.​
​
sed -i '/LimitNOFILE=10xxx76/a LimitMEMLOCK=infinity\nLimitSTACK=infinity' /lib/systemd/system/containerd.service && systemctl daemon-reload && systemctl restart containerd​
​
​​​
​
​
​
Last updated: 2026.07.23 14:31:47