commit 55f309acae2b1032b4c28f343fac9347e9251494
author Ryan <rsavitski@google.com> Tue Apr 13 18:20:53 2021 +0100
committer Ryan <rsavitski@google.com> Tue Apr 13 18:20:53 2021 +0100
tree 9bacdef812d50551f9ea6feae8a42e3793326c95
parent 016efc5e9a76abec192c26ecc6dda4e9b25a3ec6

traceconv: produce less redundant pprofs, at the expense of more work

Additionally, add an option to suffix the frame name with the
annotation, as obtained via experimental_annotated_callstack. This is
primarily relevant to Android profiles. This is enabled by default
when invoking as traceconv||trace_to_text, and suppressed if passed
"--no-annotations" flag.

The primary change for pprof_builder is that it now interns all
callsites and frames, using their contents for the identity (instead of
relying on row ids from SQL tables). This avoids emitting redundant
functions. One reason when it'd happen: stack_profile_frame has rel_pc
as part of the row's identity, whereas in the profile it is part of the
location (and therefore all such frames should be shared). Another cause
of redundancy is incremental state clears.

This does come at a cost of both more work, and a higher memory
footprint, as we're now keeping all of the stack_profile_frame and
stack_profile_callsite derived information alive in the interning
structures. I'm assuming we're not hyper-sensitive to these regressions,
as long as WASM builds don't go oom (and there I'd expect the
trace_processor state to be more significant anyway).

I've tried it on one traced_perf field trace (with ~10 sub-profiles
written, one per pid). Runtime went from ~1s to ~1.5s. Combined pprof
size went from ~600 KB to ~400 KB. Do tell if you want concrete memory
before/after numbers as well (but I assume they're on the order of the
serialized profile's size).

Change-Id: Id7d903003d31427b8b228ec34a123ab15eeb001a