Hi @YEH
Since you only mentioned "ECMP" I would make the presumption that you ask about load sharing over routing paths (L3 ECMP), and not over links inside a port-channel.
The very first important thing you must know about Nexus 9000, is that all the forwarding (L3 forwarding or L2 forwarding) will always happen in hardware. This means automatically low latency and this is why the N9K is marketed as Data Center switches.
On Nexus 9000, there is this forwarding architecture:
1. Any routing protocol will determine its best routes based on the metrics configured for the protocol and installs up to the protocol-configured maximum paths in the unicast RIB. T
2. The unicast RIB compares the administrative distances of all routing protocol paths in the RIB and selects a best path set from all of the path sets installed by the routing protocols.
3. The unicast RIB installs this best path set into the FIB for use by the forwarding plane.
The forwarding plane uses a load-sharing algorithm to select one of the installed paths in the FIB to use for a given data packet.
Load sharing uses the same path for all packets in a given flow. A flow is defined by the load-sharing method that you configure. For example, if you configure source-destination load sharing, then all packets with the same source IP address and destination IP address pair follow the same path.
You can configure the unicast FIB load-sharing algorithm using this command:
ip load-sharing address { destination port destination | source-destination [ port source-destination ]}
To verify the global load-sharing configuration, or even the hash for a specific flow, you can use these commands:
show ip load-sharing
show routing hash source-addr dest-addr
