Profiler pprof encoding benchmark and first improvements

benchmarks/profiler_submission.rb

@@ -0,0 +1,90 @@
+require 'benchmark/ips'
+require 'ddtrace'
+require 'pry'
+
+# This benchmark measures the performance of encoding pprofs and trying to submit them
+#
+# The FLUSH_DUMP_FILE (by default benchmarks/data/profiler-submission-marshal.gz, gathered from benchmarking using
+# the discourse forum rails app) can be generated by changing the scheduler.rb#flush_events to dump the contents of
+# "flush" to a file during a benchmark execution:
+#
+# dump_file = "marshal-#{Time.now.utc.to_i}.dump"
+# File.open(dump_file, "w") { |f| Marshal.dump(flush, f) }
+# Datadog.logger.info("Dumped to #{dump_file}")
+#
+# And then gzipping the result. (This can probably be automated a bit by adding an extra exporter, but the above worked
+# for me).
+
+class ProfilerSubmission
+  def create_profiler
+    @adapter_buffer = []
+
+    Datadog.configure do |c|
+      # c.diagnostics.debug = true
+      c.profiling.enabled = true
+      c.tracer.transport_options = proc { |t| t.adapter :test, @adapter_buffer }
+    end
+
+    # Stop background threads
+    Datadog.profiler.shutdown!
+
+    # Call exporter directly
+    @exporter = Datadog.profiler.scheduler.exporters.first
+    @flush = Marshal.load(
+      Zlib::GzipReader.new(File.open(ENV['FLUSH_DUMP_FILE'] || 'benchmarks/data/profiler-submission-marshal.gz'))
+    )
+  end
+
+  def check_valid_pprof
+    output_pprof = @adapter_buffer.last[:form]["data[0]"].io
+
+    expected_hashes = [
+      "75c65dec2d5d750faca3e905486c39ba44219636ac1bea15d5d624b741d4b62a",
+      "6780c47b3e271f2abe0346fd40ac358f6a34270b3a8f5743aacd1970bbbbc6f5"
+    ]
+    current_hash = Digest::SHA256.hexdigest(Zlib::GzipReader.new(output_pprof).read)
+
+    if expected_hashes.include?(current_hash)
+      puts "Output hash #{current_hash} matches known signature"
+    else
+      puts "WARNING: Unexpected pprof output -- unknown hash. Hashes seem to differ due to some of our dependencies changing, " \
+        "but it can also indicate that encoding output has become corrupted."
+    end
+  end
+
+  def run_benchmark
+    Benchmark.ips do |x|
+      x.config(time: 10, warmup: 2)
+
+      x.report("exporter #{ENV['CONFIG']}") do
+        run_once
+      end
+
+      x.save! 'profiler-submission-results.ipsbench'
+      x.compare!
+    end
+  end
+
+  def run_forever
+    run_once while true
+  end
+
+  def run_once
+    @adapter_buffer.clear
+    @exporter.export(@flush)
+  end
+end
+
+puts "Current pid is #{Process.pid}"
+
+ProfilerSubmission.new.instance_exec do
+  create_profiler
+  run_once
+  check_valid_pprof
+
+  if ARGV.include?('--forever')
+    run_forever
+  else
+    run_benchmark
+  end
+end

lib/ddtrace/profiling/pprof/builder.rb

@@ -26,6 +26,9 @@ def initialize
           @sample_types = MessageSet.new
           @samples = []
           @string_table = StringTable.new
+          # Cache these procs, since it's pretty expensive to keep recreating them
+          @build_location = method(:build_location).to_proc
+          @build_function = method(:build_function).to_proc
         end
 
         def encode_profile(profile)
@@ -60,7 +63,7 @@ def build_locations(backtrace_locations, length)
               # Function name
               backtrace_location.base_label,
               # Build function
-              &method(:build_location)
+              &@build_location
             )
           end
 
@@ -73,7 +76,7 @@ def build_locations(backtrace_locations, length)
               ''.freeze,
               0,
               "#{omitted} #{desc}",
-              &method(:build_location)
+              &@build_location
             )
           end
 
@@ -87,7 +90,7 @@ def build_location(id, filename, line_number, function_name = nil)
               @functions.fetch(
                 filename,
                 function_name,
-                &method(:build_function)
+                &@build_function
               ).id,
               line_number
             )]