Are there network switches under your control? It sounds similar to what happens when MTU on the systems MTU do not match or one system MTU is set above the value on the switch ports.
Next time the issue occurs use ping with the do not fragment flag.
ex $ ping -m DO -s 8972 ip.address
This example should be the highest value to work in the case of MTU size 9000, there is 28 byte overhead for IPv4 packets.
Second, are you sure no one is attaching to the network and duplicating the MAC address of your NFS server or perhaps the system that is stalled? If the switches are manageable you would have to insure that the MAC addresses are being learned on the correct ports.
-Jamie