Linux – Do I need root (admin) permissions to run userspace ‘perf’ tool? (perf events are enabled in Linux kernel)

How to install perf userland tool as non-root

1. Get/find sources for kernel-2.6.36-gentoo-r4 (in Gentoo Linux). The first check from this answer

   Actually, first you should look at /usr/src/linux and see if the kernel sources are still installed. You could just copy them to a directory you can write to.)

   was enough, though instead of copying whole kernel sources I just linked them:

   $ mkdir -p build
   $ cd build
   $ ln -s /usr/src/linux-2.6.36-gentoo-r4
   

2. Create directory where perf would be built, as I won't be able to write in ~/build/linux-2.6.36-gentoo-r4 directory.

   $ mkdir -p perf
   

   ^{Actually it was not what I did at first... error messages from make were entirely unhelpful at first.}

3. Go to tools/perf directory in kernel sources

   $ cd linux-2.6.36-gentoo-r4/tools/perf
   

4. Build perf, not forgetting about passing O=<destdir> option to makefile as the directory is not writable (there would be no such problem if I copied rather than symlinked kernel sources).

   $ make O=~/build/perf -k 
   Makefile:565: newt not found, disables TUI support. Please install newt-devel or libnewt-dev
       * new build flags or prefix
       CC ~/build/perf/perf.o
       CC ~/build/perf/builtin-annotate.o
       [...]
       CC ~/build/perf/util/scripting-engines/trace-event-python.o
       CC ~/build/perf/scripts/python/Perf-Trace-Util/Context.o
       AR ~/build/perf/libperf.a
       LINK ~/build/perf/perf
   ~/build/perf/libperf.a(trace-event-perl.o): In function `define_flag_value':
   ~/build/linux-2.6.36-gentoo-r4/tools/perf/util/scripting-engines/trace-event-perl.c:127: undefined reference to `PL_stack_sp'
   ~/build/linux-2.6.36-gentoo-r4/tools/perf/util/scripting-engines/trace-event-perl.c:131: undefined reference to `Perl_push_scope'
   [...]
   ~/build/perf/libperf.a(trace-event-python.o): In function `handler_call_die':
   ~/build/linux-2.6.36-gentoo-r4/tools/perf/util/scripting-engines/trace-event-python.c:53: undefined reference to `PyErr_Print'
   [...]
   collect2: ld returned 1 exit status
   make: *** [/home/narebski/build/perf/perf] Error 1
       GEN perf-archive
   make: Target `all' not remade because of errors.
   

5. Google for "undefined reference to `Perl_push_scope'". Find Fail to install perf on slackware 13.1 on unix.stackexchange.com. Follow the advice in self answer, or to be more excat the diagnosis:

   $ make O=~/build/perf -k NO_LIBPERL=1 NO_LIBPYTHON=1
   Makefile:565: newt not found, disables TUI support. Please install newt-devel or libnewt-dev
       * new build flags or prefix
       CC ~/build/perf/perf.o
       CC ~/build/perf/builtin-annotate.o
       [...]
       CC ~/build/perf/util/probe-finder.o
       AR ~/build/perf/libperf.a
       LINK ~/build/perf/perf
       GEN perf-archive
   

   Note that it is workaround rather than a solution (I have libperl.so).

6. Check Makefile for default install destination: its $(HOME). Install perf in one's own home directory:

   $ make O=~/build/perf -k NO_LIBPERL=1 NO_LIBPYTHON=1 install
   Makefile:565: newt not found, disables TUI support. Please install newt-devel or libnewt-dev
       GEN perf-archive
   install -d -m 755 '~/bin'
   install ~/build/perf/perf '~/bin'
   [...]
   install scripts/python/bin/* -t '~/libexec/perf-core/scripts/python/bin'
   

7. Check that ~/bin is in PATH

8. Check that perf works correctly (don't forget to cd in writable directory):

   $ cd
   $ perf record -f -- sleep 10
   [ perf record: Woken up 1 times to write data ]
   [ perf record: Captured and wrote 0.001 MB perf.data (~61 samples) ]
   

The output is a bit redacted, replacing my home directory with ~.

If you are distributing the computations with MPI, then using an MPI-aware tool would give you more sensible results: with a distributed application, you might have issues of load imbalance, where one MPI process is idle waiting for data to come from other processes. If you happen to be profiling exactly that MPI process, your performance profile will be all wrong.

So, the first step is usually to find out about the communication and load balance pattern of your program, and identify a sample input that gives you the workload you want (e.g., CPU-intensive on rank 0) For instance, mpiP is an MPI profiling tool that can produce a very complete report about the communication pattern, how much time each MPI call took, etc.

Then you can run a code profiling tool on one or more selected MPI ranks. Anyway, using perf on a single MPI rank is likely not a good idea because its measurements will contain also the performance of the MPI library code, which is probably not what you are looking for.