(Investigation) Multi-process support in tracy

[Arpafaucon]

This is not a bug report, but rather an attempt at documenting my work on #471.

I am still not experienced in the tracy code base, so I'd like to document a bit closely what I try, and what it gave.

• worst case scenario: it disturbs no one, and it forces me to document clearly where I'm at
• best case: more experienced people have the time to read this and can give me a few pointers to help me progress / explain what's unclear

What happened today

• Investigative work to better understand the structure. Key takeaways
  • Worker is a very central class that manages all the relevant data about a profiling run
  • to save memory space, tracy does a lot of internal data "compression"
    • thread IDs are reduced to uint16_t
    • strings are interned and retrieved by a unique ID
• Some interesting readings
• export-csv gave me good insights on the important internal data structures of Worker
• and import-chrome on the other side shows how to feed data into a Worker

I tried testing #766 but I had no success so far, I get the error reported in #766 (comment)

To get my hands dirty, I explored an alternative approach that only supports offline files, let's call it tracy-merge.

• we generate multiple profiler runs separately (they talk to different tracy-capture processes)
• we run tracy-merge on the set of generated *.tracy, to generate one file that contains all the events
• to limit complexity, I use the importing constructor of Worker. Basically I fill a big vector with events of all the profiling files in turn, taking care of re-mapping thread IDs to avoid overlaps

Current result, generated from merging the same profiling run from test/test_cpp copied twice.

• I lose the real thread IDs in the merge process
• I still have missing zone names (but some of them are present) - I can't explain this at the moment
• I only converted timeline events for now, so plots and frames are lost

[Arpafaucon]

Report

I got back on the multiplexer work, trying to figure out the reason of the failure. What is still confusing is that the experimental profiler (as I read it) seems to want to consider all client-issued QueueItems as answers to server queries.

• adding some logic to skip the handle_client_response for the non-response frames partly fixes the issue, and i can now spawn the multpilexer with multiple clients. However, I have recurrent crashes to investigate
• I was kinda scared by the huge switch case in get_time_and_field that is needed to remap the time correctly (still have issues there), but I agree with cipharius: not a lot we can do without refactoring the profiler and/or the worker
• discovered with pleasure TracyEventDebug, very nice to understand the protocol by example
• out of topic: merged [IDE] migrate test folder to CMake configuration #824
• useful resources
  • the discord channel
  • https://wolf.nereid.pl/posts/tracy-internals (and the other blog post about tracy as well)
  • the source code (spent hours in it)