NCCL 算法 & 参数选择:
pick `Ring+Simple` for large messages and `Tree+LL128` for small ones.
default autotune may suffice.
---
<\/li>
c
---
二、多核并行网络层面细节解析及常见错误修复思路
# QPS & Rail 调优 :<\/b>
prompt 如:
fqps=`export NCCL_IB_QPS_PER_CONNECTION=` &&
enable_split=`export NCCL_IB_SPLIT_DATA_ON_QPS=`
and check link utilization.
but if you see *unbalanced traffic* in per‑rank logs n increase QPs.
e.g.,set QPs from default to *32* and observe busbw rise by ~十五成上下.
从note来看。this reduces hash collision on ECMP routers.
breaking point:
default single‑rail QP may cause severe contention when scaling beyond 64 GPUs per rack.\t—\t \t\t\t\t\t\t\t\t\t \t \t \t \t \t \t \t \t . ...\r // ... ... ... ... ..\r // .................... ....... . .......... .......... .................
三、大模型训练网络故障域与监控常用方法
*fail-stop vs fail-slow* :<\/b>
live monitoring of NIC health via SNMP or vendor APIs can catch early degradation.
always compare per‑rank timings – a sudden jump indicates slow link before it fails completely.