ebpf: aggregate python stacks in ebpf program instead of userspace

[korniltsev]

We used to stream each sample once it is captured, now we increment counter in ebpf world, and dump it once every profiling interval.

This seem to consume less cpu

Before:
         PM   UID       PID    %usr %system  %guest   %wait    %CPU   CPU  Command
Average:        0    382723    4.57    0.88    0.00    0.00    5.45     -  ___ebpf_python_

After
         PM   UID       PID    %usr %system  %guest   %wait    %CPU   CPU  Command
Average:        0    803223    2.43    0.18    0.00    0.00    2.62     -  ___ebpf_python_

[simonswine]

LGTM

I noticed a small diff comparing the old python profiles with the new ones. Can't really tell if this is something significant wrong. Maybe just have a quick look:

https://flamegraph.com/share/2cd304bb-cf01-11ee-a059-7a6c3496f108

Green is the profile collected using #2996 and red is before

[korniltsev]

I noticed a small diff comparing the old python profiles with the new ones. Can't really tell if this is something significant wrong. Maybe just have a quick look:

https://flamegraph.com/share/2cd304bb-cf01-11ee-a059-7a6c3496f108

Green is the profile collected using #2996 and red is before

I've also added kernel traces in this PR and it looks like the only difference is that new version has kernel stacks in places of
fibonacci.py _get_from_memoization etc.

Does it look reasonable?

[simonswine]

Does it look reasonable?

Yes makes total sense, missed the kernel stacks change tbf 🙂