stackprof

Stackprof

A sampling call-stack profiler for Ruby.

Inspired heavily by gperftools, and written as a replacement for perftools.rb.

Requirements

• Ruby 2.2+
• Linux-based OS

Getting Started

Install

In your Gemfile add:

gem 'stackprof'

Then run $ bundle install. Alternatively you can run $ gem install stackprof.

Run

in ruby:

StackProf.run(mode: :cpu, out: 'tmp/stackprof-cpu-myapp.dump') do
  #...
end

via rack:

use StackProf::Middleware, enabled: true,
                           mode: :cpu,
                           interval: 1000,
                           save_every: 5

reporting:

…

For an experimental version of WebUI reporting of stackprof, see stackprof-webnav

To generate flamegraphs with Stackprof, additional data must be collected using the raw: true flag. Once you've collected results with this flag enabled, generate a flamegraph with:

$ stackprof --flamegraph tmp/stackprof-cpu-myapp.dump > tmp/flamegraph

After the flamegraph has been generated, you can generate a viewer command with:

$ stackprof --flamegraph-viewer=tmp/flamegraph

The --flamegraph-viewer command will output the exact shell command you need to run in order to open the tmp/flamegraph you generated with the built-in stackprof flamegraph viewer:

Alternatively, you can generate a flamegraph that uses d3-flame-graph:

$ stackprof --d3-flamegraph tmp/stackprof-cpu-myapp.dump > flamegraph.html

And just open the result by your browser.

Sampling

Four sampling modes are supported:

• :wall (using ITIMER_REAL and SIGALRM) [default mode]
• :cpu (using ITIMER_PROF and SIGPROF)
• :object (using RUBY_INTERNAL_EVENT_NEWOBJ)
• :custom (user-defined via StackProf.sample)

Samplers have a tuneable interval which can be used to reduce overhead or increase granularity:

• Wall time: sample every interval microseconds of wallclock time (default: 1000)

StackProf.run(mode: :wall, out: 'tmp/stackprof.dump', interval: 1000) do
  #...
end

• CPU time: sample every interval microseconds of CPU activity (default: 1000 = 1 millisecond)

StackProf.run(mode: :cpu, out: 'tmp/stackprof.dump', interval: 1000) do
  #...
end

• Object allocation: sample every interval allocations (default: 1)

StackProf.run(mode: :object, out: 'tmp/stackprof.dump', interval: 1) do
  #...
end

By default, samples taken during garbage collection will show as garbage collection frames including both mark and sweep phases. For longer traces, these can leave gaps in a flamegraph that are hard to follow. They can be disabled by setting the ignore_gc option to true. Garbage collection time will still be present in the profile but not explicitly marked with its own frame.

Samples are taken using a combination of three new C-APIs in ruby 2.1:

• Signal handlers enqueue a sampling job using rb_postponed_job_register_one. this ensures callstack samples can be taken safely, in case the VM is garbage collecting or in some other inconsistent state during the interruption.

• Stack frames are collected via rb_profile_frames, which provides low-overhead C-API access to the VM's call stack. No object allocations occur in this path, allowing stackprof to collect callstacks in allocation mode.

• In allocation mode, samples are taken via rb_tracepoint_new(RUBY_INTERNAL_EVENT_NEWOBJ), which provides a notification every time the VM allocates a new object.

Aggregation

Each sample consists of N stack frames, where a frame looks something like MyClass#method or block in MySingleton.method. For each of these frames in the sample, the profiler collects a few pieces of metadata:

• samples: Number of samples where this was the topmost frame
• total_samples: Samples where this frame was in the stack
• lines: Samples per line number in this frame
• edges: Samples per callee frame (methods invoked by this frame)

The aggregation algorithm is roughly equivalent to the following pseudo code:

trap('PROF') do
  top, *rest = caller

  top.samples += 1
  top.lines[top.lineno] += 1
  top.total_samples += 1

  prev = top
  rest.each do |frame|
    frame.edges[prev] += 1
    frame.total_samples += 1
    prev = frame
  end
end

This technique builds up an incremental call graph from the samples. On any given frame, the sum of the outbound edge weights is equal to total samples collected on that frame (frame.total_samples == frame.edges.values.sum).

Reporting

Multiple reporting modes are supported:

• Text
• Dotgraph
• Source annotation

StackProf::Report.new(data).print_text

     TOTAL    (pct)     SAMPLES    (pct)     FRAME
        91  (48.4%)          91  (48.4%)     A#pow
        58  (30.9%)          58  (30.9%)     A.newobj
        34  (18.1%)          34  (18.1%)     block in A#math
       188 (100.0%)           3   (1.6%)     block (2 levels) in <main>
       185  (98.4%)           1   (0.5%)     A#initialize
        35  (18.6%)           1   (0.5%)     A#math
       188 (100.0%)           0   (0.0%)     <main>
       188 (100.0%)           0   (0.0%)     block in <main>
       188 (100.0%)           0   (0.0%)     <main>

StackProf::Report.new(data).print_graphviz

…

StackProf::Report.new(data).print_method(/pow|newobj|math/)

…

Usage

The profiler is compiled as a C-extension and exposes a simple api: StackProf.run(mode: [:cpu|:wall|:object]). The run method takes a block of code and returns a profile as a simple hash.

# sample after every 1ms of cpu activity
profile = StackProf.run(mode: :cpu, interval: 1000) do
  MyCode.execute
end

This profile data structure is part of the public API, and is intended to be saved (as json/marshal for example) for later processing. The reports above can be generated by passing this structure into StackProf::Report.new.

The format itself is very simple. It contains a header and a list of frames. Each frame has a unique ID and identifying information such as its name, file, and line. The frame also contains sampling data, including per-line samples, and a list of relationships to other frames represented as weighted edges.

{:version=>1.0,
 :mode=>:cpu,
 :inteval=>1000,
 :samples=>188,
 :missed_samples=>0,
 :frames=>
  {70346498324780=>
    {:name=>"A#pow",
     :file=>"/Users/tmm1/code/stackprof/sample.rb",
     :line=>11,
     :total_samples=>91,
     :samples=>91,
     :lines=>{12=>91}},
   70346498324900=>
    {:name=>"A#initialize",
     :file=>"/Users/tmm1/code/stackprof/sample.rb",
     :line=>5,
     :total_samples=>185,
     :samples=>1,
     :edges=>{70346498324780=>91, 70346498324680=>58, 70346498324580=>35},
     :lines=>{8=>1}},

Above, A#pow was involved in 91 samples, and in all cases it was at the top of the stack on line 12.

A#initialize was in 185 samples, but it was at the top of the stack in only 1 sample. The rest of the samples are divided up between its callee edges. All 91 calls to A#pow came from A#initialize, as seen by the edge numbered 70346498324780.

Advanced usage

The profiler can be started and stopped manually. Results are accumulated until retrieval, across multiple start/stop invocations.

StackProf.running? # => false
StackProf.start(mode: :cpu)
StackProf.running? # => true
StackProf.stop
StackProf.results('/tmp/some.file')

All options

StackProf.run accepts an options hash. Currently, the following options are recognized:

Todo

• file/iseq blacklist
• restore signal handlers on stop