【NVIDIA-H100-UFM故障分析实战】GPU 节点反复 “掉线又上线“？300 条日志揪出 InfiniBand 链路抖动真凶

一、日志核心信息提取与现象全景分析

#报错日志
Jul 01 08:01:17 549674 [92ADE640] 0x02 -> drop_mgr_remove_port: Removed port with GUID:asdfas923782sjs LID range [156, 156] of node:gpu015 mlx5_8
Jul 01 08:01:24 495087 [92ADE640] 0x02 -> state_mgr_report_new_ports: Discovered new port with GUID:asdfas923782sjs LID range [156,156] of node: gpu015 mlx5_8
Jul 01 09:04:10 490319 [92ADE640] 0x02 -> drop_mgr_remove_port: Removed port with GUID:asdfas923782sjs LID range [156, 156] of node:gpu015 mlx5_8
Jul 01 09:04:15 942861 [92ADE640] 0x02 -> state_mgr_report_new_ports: Discovered new port with GUID:asdfas923782sjs LID range [156,156] of node: gpu015 mlx5_8
Jul 01 15:45:45 187515 [92ADE640] 0x02 -> drop_mgr_remove_port: Removed port with GUID:asdfas923782sjs LID range [156