Nuts and bolts Talk.
Debugging TCP session is tricky. While looking at aggregate TCP stats may help in some cases it falls short for use cases which require discovering and fixing latency issues in highly distributed applications. To cover such a use case, one would need to collect the kernel stack's point of view for the sessions lifetime.
Merely using TCP_INFO doesnt cut it for use cases when we need to see more fine grained details for each message such as is the case for structured RPC request and response interactions.
In this talk, Soheil Hassas Yeganeh and Yuchung Cheng will show how to address these issues using timestamps to track individual messages in the kernel. As a crucial extension, they will show how we to accurately capture the state of TCP using OPT_STATS for individual messages. Further, they enumerate challenges, shortcomings and gotchas of kernel timestamps when deployed at scale.
More info: https://netdevconf.org/0x13/session.html?talk-tcp-timestamping
A reminder to all: Early bird registration is still open until Feb 20th. https://netdevconf.org/0x13/registration.html
cheers, jamal