Blackbird - SB:Switch Event Latency

Reference:

• for Test Plan see: Test Plan - Perf & Scale-out;  
• ONLabTest script - https://github.com/opennetworkinglab/ONLabTest: TopoPerfNextBM

Test Setup:

• RC2 - commit 4964d3d768deb0a8b5773c84f47787db455ceaba
• Using OVS version 2.3.1/OF 1.3
• Bare-metal Servers: dual-Xeon E5-2670, 32GB DDR3 RAM, SSD, Cluster Network is 1Gbps
• JAVA_OPTS="${JAVA_OPTS:--Xms8G -Xmx8G}"
• "org.onosproject.net.topology.impl.DefaultTopologyProvider.cfg": maxEvents = 1; maxIdleMs = 0; maxBatchMs = 0

Test Procedure:

• 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:

Result Discussions:

1. 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.
2. 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):
   1. TCP syn -> OFP Hello (from ovs): 0.3ms
   2. OFP Hello (from ovs) -> OFP of_features_request:  2.6ms
   3. OFP of_features_request -> OFP of_features_reply: 47.0ms
   4. OFP of_feature_reply -> OFP role_request:  8.0ms
   5. OFP role_request -> OFP role_reply: 0.2ms