GMT measurements for network traffic are not representative for HTTP

[ArneTR]

Currently GMT measures the network traffic directly from the network adapter.

This means it polls values from procfs by stepping into the relevant namespace.

On the host this typically leads to the fact of including traffic unrelated to the application benchmarked (which is by design) (See: #1322)

In the cgroup namespace it is fully isolated to the application, but is still on the byte level transferred. This means it will also include retransmitted packages and overhead that is not coming from the TCP/UDP payloads.

It would be nice to have functionality to get the actual transmitted HTTP traffic. Similar to what the Chrome DevTools network console shows.

At the moment there is no clear idea how to best do that.
A possible approach would be to use TinyProxy, which we have as a metrics provider, and use the information of the HTTP request size.
However the time resolution here is extremely bad. Unclear is also how it behaves for streaming UDP and / or QUIC.

This issue is an open call for:

• recommendations
• design ideas
• questions
• and also to probe interest in this feature

At the moment we feel that it is sufficient to have the network data AS IT REALLY HAPPENED. Having HTTP would be nice, but is not necessarily needed to profile an application.

Happy for any comments :)

[ArneTR]

@davidkopp