Elixir Support

[dogukanzorlu]

Hi Parca peoples,

for Parca to provide Elixir support, I developed a library that imitates net/http/pprof in golang and generates meaningful pprof profile data from internal erlang library :fprof.

Many explanations you need are available in the documentation. But still, I will write the same ones and the additions here.

Aim The observation and profiling culture of erlang and elixir is different from other languages. Elixir/Erlang runs on a virtual machine (a.k.a beam) and each block of code represents a process. Erlang provides many tools internally for monitoring abstracted processes and processes dependent functions. This library produces as meaningful pprof output as possible using the built-in Erlang profiling tools. Thus, it tries to spread the Erlang/Elixir observability with pprof-based ad-hoc profiling tools.

A few important things

• This library an experimental and still under development, a library that needs to be careful to use on the production line.
• The accuracy of the outputs has been proven by testing with cross-tools. But this alpha version does not offer a full pprof service.
• fprof significantly slows down the application it is running on. It monitors all processes and collects their tracing data. Therefore, even a 10-second data collection in large-scale applications can mean GBs of data. It is recommended that the scrape seconds do not exceed 5 in order not to get lost in the abundance of data. More information: :fprof

• The :suspend pseudo function has got an OWN time of zero. This is to prevent the process total OWN time from including time in suspension. Whether suspend time is really ACCUMULATED or OWN time is more of a philosophical question. more: analysis-format

  Installation and Usage

  Installation

  def deps do
   [
    {:pprof, "~> 0.1.0"}
   ]
  end

  After:

  $ mix deps.get

  Usage

  Add pprof config your config file like this:

  config :pprof, :port, 8080

  You can use with go pprof:

  $ go tool pprof http://localhost:8080/debug/pprof/profile?seconds=5&type=fprof
  

  Also using with Parca add configure in parca.yaml:

      # params:
      #   type: [ 'fprof' ]

  

source: github Thanks! @kakkoyun @brancz

[kakkoyun]

xref https://github.com/parca-dev/parca-agent/issues/145

[brancz]

Very cool! I'm curious, what do the numbers before the closing curly brackets mean are they parameters? For a 5 seconds profile 1.1k samples seems like a lot, what's the sampling frequency?

[dogukanzorlu]

{ModuleName, Function, Arity} so its mean for example Erlang.garbage_collect(arg1, arg2) -> {:erlang, :garbage_collect, 2}. actually fprof is not simple call stack sampling, fprof traces the entire process on the beam and the invoked functions that depend on the processes. therefore most of time sampling so huge. Unfortunately, I couldn't find any content about frequency in fprof and its source code. @brancz

[brancz]

{ModuleName, Function, Arity} so its mean for example Erlang.garbage_collect(arg1, arg2) -> {:erlang, :garbage_collect, 2}.

Got it! Thanks for the explanation!

actually fprof is not simple call stack sampling, fprof traces the entire process on the beam and the invoked functions that depend on the processes. therefore most of time sampling so huge. Unfortunately, I couldn't find any content about frequency in fprof and its source code.

Judging by the docs I would be careful with using this continuously, it seems like fprof and eprof are meant as one-off tools (even the docs claim they add significant overhead). Only cprof seems to be sampling and <10% impact (still rather large for continuous profiling purposes).

I would suggest doing the CPU profiling with the Parca Agent and using the ERL_FLAGS="+S 1 +JPperf true" flag on the erlang/elixir application. This should give great visibility at very low overhead.

Anything memory-related, especially heap is something that would make a ton of sense to be in a library like you created! In any case, nice work looking forward to seeing more!

[dogukanzorlu]

Exactly you are right about continuously. But considering that there is no pprof in elixir or erlang, it's not bad for a start. I will also produce pprof for memory and process. I am also working on perf maps related VM. In any case, thank you for your comments. @brancz

[brancz]

Thank you so much for this work, the community needs this type of engagement!