Profiling the Tezos node

Memory profiling the OCaml heap

  • Install an OCaml switch with the statmemprof patch:

    4.04.2+statistical-memprof or 4.06.0+statistical-memprof

  • Install statmemprof-emacs.

  • Enable loading statmemprof into the node.

    Add the statmemprof-emacs package as a dependency to the main package, and add let () = Statmemprof_emacs.start 1E-4 30 5 to the node_main.ml file.

    Arguments:

    • sampling_rate is the sampling rate of the profiler. Good value: 1e-4.

    • callstack_size is the size of the fragment of the call stack which is captured for each sampled allocation.

    • min_sample_print is the minimum number of samples under which the location of an allocation is not displayed.

  • Load sturgeon into emacs, by adding this to your .emacs:

(let ((opam-share (ignore-errors (car (process-lines "opam" "config" "var" "share")))))
 (when (and opam-share (file-directory-p opam-share))
  (add-to-list 'load-path (expand-file-name "emacs/site-lisp" opam-share))))

(require 'sturgeon)
  • Launch the node then connect to it with sturgeon.

    If the process is launched with pid 1234 then

  M-x sturgeon-connect
  tezos-nodememprof.1234.sturgeon

(tab-completion works for finding the socket name)

Memory profiling the C heap

  • Install valgrind and massif-visualizer

valgrind --tool=massif tezos-node run ...
  • Stop with Ctrl-C then display with

massif-visualizer massif.out.pid

Performance profiling

  • Install perf (the linux-perf package for debian).

    If the package does not exist for your current kernel, a previous version can be used. Substitute the perf command to perf_4.9 if your kernel is 4.9).

  • Either:

    • Run the node, find the pid.

      Attach perf with perf record -p pid -F 99 --call-stack dwarf.

      Then stop capturing with Ctrl-C. This can represent a lot of data. Don’t do that for too long. If this is too much you can remove the --call-stack dwarf to get something more manageable, but interpreting the information can be harder.

    • Let perf run tezos-node: perf record -g -F 99 --call-graph=dwarf -- ./tezos-node run ...

      This will write perf.data after having stopped the node with Ctrl-C.

    In both cases, the -F argument specifies the frequency of sampling of data (in hertz). If too much data is generated, use a smaller value. If data is not precise enough, try using a higher value.

  • display the result with perf report, or use a more advanced visualizer (recommended). Such visualizers include: