diff options
author | Jari Vetoniemi <jari.vetoniemi@indooratlas.com> | 2020-03-16 18:49:26 +0900 |
---|---|---|
committer | Jari Vetoniemi <jari.vetoniemi@indooratlas.com> | 2020-03-30 00:39:06 +0900 |
commit | fcbf63e62c627deae76c1b8cb8c0876c536ed811 (patch) | |
tree | 64cb17de3f41a2b6fef2368028fbd00349946994 /jni/ruby/lib/profiler.rb |
Fresh start
Diffstat (limited to 'jni/ruby/lib/profiler.rb')
-rw-r--r-- | jni/ruby/lib/profiler.rb | 148 |
1 files changed, 148 insertions, 0 deletions
diff --git a/jni/ruby/lib/profiler.rb b/jni/ruby/lib/profiler.rb new file mode 100644 index 0000000..e53951c --- /dev/null +++ b/jni/ruby/lib/profiler.rb @@ -0,0 +1,148 @@ +# Profile provides a way to Profile your Ruby application. +# +# Profiling your program is a way of determining which methods are called and +# how long each method takes to complete. This way you can detect which +# methods are possible bottlenecks. +# +# Profiling your program will slow down your execution time considerably, +# so activate it only when you need it. Don't confuse benchmarking with +# profiling. +# +# There are two ways to activate Profiling: +# +# == Command line +# +# Run your Ruby script with <code>-rprofile</code>: +# +# ruby -rprofile example.rb +# +# If you're profiling an executable in your <code>$PATH</code> you can use +# <code>ruby -S</code>: +# +# ruby -rprofile -S some_executable +# +# == From code +# +# Just require 'profile': +# +# require 'profile' +# +# def slow_method +# 5000.times do +# 9999999999999999*999999999 +# end +# end +# +# def fast_method +# 5000.times do +# 9999999999999999+999999999 +# end +# end +# +# slow_method +# fast_method +# +# The output in both cases is a report when the execution is over: +# +# ruby -rprofile example.rb +# +# % cumulative self self total +# time seconds seconds calls ms/call ms/call name +# 68.42 0.13 0.13 2 65.00 95.00 Integer#times +# 15.79 0.16 0.03 5000 0.01 0.01 Fixnum#* +# 15.79 0.19 0.03 5000 0.01 0.01 Fixnum#+ +# 0.00 0.19 0.00 2 0.00 0.00 IO#set_encoding +# 0.00 0.19 0.00 1 0.00 100.00 Object#slow_method +# 0.00 0.19 0.00 2 0.00 0.00 Module#method_added +# 0.00 0.19 0.00 1 0.00 90.00 Object#fast_method +# 0.00 0.19 0.00 1 0.00 190.00 #toplevel + +module Profiler__ + class Wrapper < Struct.new(:defined_class, :method_id, :hash) # :nodoc: + private :defined_class=, :method_id=, :hash= + + def initialize(klass, mid) + super(klass, mid, nil) + self.hash = Struct.instance_method(:hash).bind(self).call + end + + def to_s + "#{defined_class.inspect}#".sub(/\A\#<Class:(.*)>#\z/, '\1.') << method_id.to_s + end + alias inspect to_s + end + + # internal values + @@start = nil # the start time that profiling began + @@stacks = nil # the map of stacks keyed by thread + @@maps = nil # the map of call data keyed by thread, class and id. Call data contains the call count, total time, + PROFILE_CALL_PROC = TracePoint.new(*%i[call c_call b_call]) {|tp| # :nodoc: + now = Process.times[0] + stack = (@@stacks[Thread.current] ||= []) + stack.push [now, 0.0] + } + PROFILE_RETURN_PROC = TracePoint.new(*%i[return c_return b_return]) {|tp| # :nodoc: + now = Process.times[0] + key = Wrapper.new(tp.defined_class, tp.method_id) + stack = (@@stacks[Thread.current] ||= []) + if tick = stack.pop + threadmap = (@@maps[Thread.current] ||= {}) + data = (threadmap[key] ||= [0, 0.0, 0.0, key]) + data[0] += 1 + cost = now - tick[0] + data[1] += cost + data[2] += cost - tick[1] + stack[-1][1] += cost if stack[-1] + end + } +module_function + # Starts the profiler. + # + # See Profiler__ for more information. + def start_profile + @@start = Process.times[0] + @@stacks = {} + @@maps = {} + PROFILE_CALL_PROC.enable + PROFILE_RETURN_PROC.enable + end + # Stops the profiler. + # + # See Profiler__ for more information. + def stop_profile + PROFILE_CALL_PROC.disable + PROFILE_RETURN_PROC.disable + end + # Outputs the results from the profiler. + # + # See Profiler__ for more information. + def print_profile(f) + stop_profile + total = Process.times[0] - @@start + if total == 0 then total = 0.01 end + totals = {} + @@maps.values.each do |threadmap| + threadmap.each do |key, data| + total_data = (totals[key] ||= [0, 0.0, 0.0, key]) + total_data[0] += data[0] + total_data[1] += data[1] + total_data[2] += data[2] + end + end + + # Maybe we should show a per thread output and a totals view? + + data = totals.values + data = data.sort_by{|x| -x[2]} + sum = 0 + f.printf " %% cumulative self self total\n" + f.printf " time seconds seconds calls ms/call ms/call name\n" + for d in data + sum += d[2] + f.printf "%6.2f %8.2f %8.2f %8d ", d[2]/total*100, sum, d[2], d[0] + f.printf "%8.2f %8.2f %s\n", d[2]*1000/d[0], d[1]*1000/d[0], d[3] + end + f.printf "%6.2f %8.2f %8.2f %8d ", 0.0, total, 0.0, 1 # ??? + f.printf "%8.2f %8.2f %s\n", 0.0, total*1000, "#toplevel" # ??? + end +end |