0

We currently use node clustering in order to get the most out of our machines and would like to be able to profile all processes simultaneously (only the function calls, we're using --perf_basic_prof_only_functions). While getting the information and building flamegraphs work fine, we seem to get a lot of entries for [perf-$PID.map] making it seem as though were either missing some invocation to tell one of the tools to account for multiple perf files.

Specifically, we're doing something similar to the following:

sudo perf record -F 99 -o perf.data -p $PIDS -g -- sleep 30

sudo perf script -i perf.data > out.nodestacks

# Using http://github.com/brendangregg/FlameGraph
./stackcollapse-perf.pl < ../out.nodestacks | ./flamegraph.pl > ../flame.svg

But looking at the output of perf script there are lots of entries similar to:

3881ddc630da [unknown] (/tmp/perf-20350.map)
3881dc5aae44 [unknown] (/tmp/perf-20350.map)
3881dc7d7275 [unknown] (/tmp/perf-20350.map)
3881dc7d6f4b [unknown] (/tmp/perf-20350.map)
3881dc7d6953 [unknown] (/tmp/perf-20350.map)

Has anyone else run into this issue? Thanks!

ttacon
  • 503
  • 3
  • 9

1 Answers1

0

Did you try with --perf_basic_prof (instead of --perf_basic_prof_only_functions)? At least that resolved for me some missing, not translated entries.

In my case this were entries like: Builtin:JSEntryTrampoline Stub:JSEntryStub

pelzelmm
  • 1
  • `--perf_basic_prof` isn't an option as we don't want to have huge continuously growing files at the moment unfortunately :( – ttacon Feb 07 '17 at 13:30