If you’re attempting to use SCP on your Nexus switch and you realize you’re getting slow performance, even with jumbo frames enabled on your source interface, the physical connected interface, and you’ve verified everything along the path is set to the correct jumbo MTU, you’re likely going to need to reference your system QOS settings for network-qos. By default, the standard policy-map applied under system qos is references the class-default network-qos class and sets the MTU to 1500. You will need to create a new policy-map like this:
policy-map type network-qos jumbo
class type network-qos class-default
service-policy type network-qos jumbo
Once you have this created, the next thing you want to do is enable path-mtu-discovery
ip tcp path-mtu-discovery
From there, you can attempt to ping your destination using a jumbo frame packet size with the df-bit set for testing, you should see it go through successfully and you’ll notice your SCP transfers are much faster for those large BIN files for code upgrades.
The smoking gun to finding this issue is going to be on your physical uplinks doing a
show interface e#/# and you’ll notice on the TX side you’ll see jumbo frames sent and then you’ll see incrementing input errors on the RX, in increasing precision with the number of jumbo frames sent.
This was testing on a Cisco Nexus 3172 running 7.0.3.I2.2 code.