switch event generate on ONOS1 by connecting ovs switch to it;
record tshark (tcp syn-ack) and graph event timestamp to calculate differences. In a multi-node test, the node with latest Graph event timestamp is used.
test is run in multiple iterations ("device-remove" is used to remove device from store between runs) to obtain statistical results
a number of warm-up runs are used before taking test.
Result:
1node
3nodes
5nodes
7nodes
Syn/Ack -> OFP role reply
62.0
51.21
74.58
73.89
OFP role reply -> topology update
3.5
5.94
4.38
2.86
Result Discussions:
Since TCP syn-ack -> OFP role_reply are related to Openflow msg exchange between OVS and ONOS, we can see that the ONOS core processing timing is the lesser portion of the total end-to-end switch up timing. The end-to-end latency stdev is in the range of 8~10ms.
Wireshark to further dissect timing reveals that most of the time spent before OFP role reply is between "of_features_request" and "of_feature_reply" - in the order of 40~60ms. Therefore, the largest latency contributor is OVS switch response time. A further breakdown of a typical timing can be found in the following example (TCP syn -> OFP role reply = 58.1ms):