WiFi Back Up, Two IPs Active | Generated by AI
Question: If WiFi was DOWN, why can I still reach the machine at 192.168.1.47?
Answer:
Because the WiFi came back up — the state DOWN in that log was just a snapshot at one moment. NetworkManager auto-reconnects, so now both NICs are live simultaneously, and the machine holds two IPs at once:
192.168.1.47→ WiFi (wlp3s0)192.168.1.28→ Ethernet
Both accept inbound SSH. This is normal multi-homing — nothing exclusive about it.
Confirm on the xiaoxin box:
ip -br addr # expect UP on both wlp3s0 and enp*, each with its IP
ip route # look at 'default via ... metric N' lines
You’ll likely see two default routes with different metrics, e.g.:
default via 192.168.1.1 dev enp3s0 metric 100 # ethernet wins (lower metric)
default via 192.168.1.1 dev wlp3s0 metric 600
That metric ordering explains the asymmetry you noticed:
- Inbound: you can connect to either IP — Linux answers on whichever interface the packet arrives, and replies keep the source IP you dialed (weak host model). So
.47works fine from your Mac. - Outbound: when the worker initiates the Gloo connection to the master at
.36, the kernel picks the default route → ethernet → master sees the worker as.28.
So: same machine, two doors. You knock on .47; the training process walks out through .28.
For training stability, make it deterministic instead of metric-dependent:
# in run_dist_cpu_worker.sh
export GLOO_SOCKET_IFNAME=enp3s0 # exact name from `ip -br link`
# and on the master (.36)
export GLOO_SOCKET_IFNAME=<its ethernet ifname>
Otherwise a WiFi flap mid-run can shift routing and hang the allreduce. Wired-pinned on both ends = stable steps.