Sub-if support

[stathis]

Hello,

It seems that sub interfaces are not supported yet like they are with native Linux ifs?

vpp# sflow enable-disable HundredGigabitEthernet3/0/0.100
sflow enable-disable: Invalid interface, only works on physical ports

The message doesn't make sense since it's supported on BondEthernet0.100 for example, which is not a physical port.

[sflow]

Thank you for testing!

The sFlow data model describes "black boxes" connected by physical ports. Sub-interface details can be sifted out at the collector because the MACs, VLANs, MPLS-Label-Stacks, tunnels etc. are always present in the feed. So you are encouraged to enable sFlow on all ports and then leave it alone. The rest of the work happens at the collector.

I acknowledge that this may be a departure from how other flow-monitoring technologies operate. It comes from the very low overhead that sFlow monitoring imposes. Enabling sFlow on every port of every device is normal and expected. It was always designed to be a "whole network" solution.

(If you browse the example configs under https://github.com/sflow/config you will see that many of them allow sFlow to be enabled on all ports with one command. We might want a simple "sflow enable all" in the VPP CLI someday, but since VPP is more likely to be configured by an "overlord" orchestration script it seemed OK to start with explicit per-interface controls.)

I haven't tested with a bond yet. Enabling on a bond should (ideally) enable sFlow on all the physical components of the bond, and then tell hsflowd what the bond if_index and MAC is for each component whenever their counters are sent. But maybe VPP does not present bonds that way? Do you get packet-samples from your bond when you configure it? See issue #13 for more details.

[stathis]

Hello,

That is indeed interesting, and since the overhead is so low, sflow enable all would make sense.

I was going with the way I've set hsflowd.conf on other machines:

...
pcap { dev=bond2.3540 }
...

Honestly I have not checked how it handles information on the bond, I should really test that since there will be cases where I would like to monitor subinterfaces on bonds.