perf overwhelms stack trace

[WallStProg]

First off -- many thanks, and especially for the work on getting combined Java/C++ stack traces. That's a big win for me, as I work a lot with C++ code hosted in a Java-based FIX engine.

The issue I'm running into is that the flame graph ends up with a single huge spike which resolves to native_write_msr_safe inside perf code -- and that completely throws off the scaling for the rest of the graph.

I tried looking for a way to exclude certain symbols from the graph, but didn't see anything. Any suggestions would be most welcome.

[WallStProg]

I turns out I have several of these, although one is particularly tall (and skinny -- almost didn't see it).

TIA!

[WallStProg]

A couple of things:

• If I chmod the perf.data file so I can run as non-root, all the native_write_msr_safe stacks disappear (or more likely get turned into [[kernel.kallsyms]]).
• I'm left with a similar problem with [[kernel.kallsyms]] , but a little judicious use of grep resolves that:

perf script | grep -v '[[kernel.kallsyms]] ' | stackcollapse-perf.pl | flamegraph.pl --color=java --hash > flamegraph.svg

Still interested in any addl insights or tips you might have...