diff --git a/.rubocop.yml b/.rubocop.yml
index f7b04c4..28ae1ea 100644
--- a/.rubocop.yml
+++ b/.rubocop.yml
@@ -5,6 +5,9 @@ inherit_gem:
 AllCops:
   TargetRubyVersion: 2.4
 
+Style/ParallelAssignment:
+  Enabled: false
+
 Metrics/MethodLength:
   Max: 15
 
diff --git a/Gemfile.lock b/Gemfile.lock
index 987c50e..cb45405 100644
--- a/Gemfile.lock
+++ b/Gemfile.lock
@@ -45,6 +45,7 @@ GEM
       unicode-display_width (>= 1.4.0, < 1.7)
     rubocop-rspec (1.33.0)
       rubocop (>= 0.60.0)
+    ruby-prof (0.18.0)
     ruby-progressbar (1.10.1)
     unicode-display_width (1.6.0)
 
@@ -56,6 +57,7 @@ DEPENDENCIES
   prometheus-client-tracer!
   pry (~> 0.10)
   rspec (~> 3.8)
+  ruby-prof (~> 0.18)
 
 BUNDLED WITH
    2.0.1
diff --git a/exe/benchmark-tracer.rb b/exe/benchmark-tracer.rb
new file mode 100755
index 0000000..db0d92a
--- /dev/null
+++ b/exe/benchmark-tracer.rb
@@ -0,0 +1,50 @@
+# frozen_string_literal: true
+
+# This file can be used to validate the performance of the tracer is not incomparible to a
+# naive tracing implementation. The performance will degrade linerarly with the number of
+# on-going traces, however a modern i7 can achieve trace throughput:
+#
+#   naive: 333k/s
+#   trace,   0 on-going: 233k/s
+#   trace,   5 on-going: 216k/s
+#   trace,  25 on-going: 155k/s
+#   trace, 100 on-going: 84k/s
+#
+# While this seems more than reasonable performance for almost any use case, consider
+# using multiple tracers if this does become a problem.
+
+require "benchmark"
+
+require "prometheus/client"
+require "prometheus/client/tracer"
+
+counter = Prometheus::Client::Counter.new(
+  :counter, docstring: "example", labels: %i[worker]
+)
+another_counter = Prometheus::Client::Counter.new(
+  :counter_another, docstring: "another example", labels: %i[]
+)
+
+def trace(metric, labels = {})
+  start = Process.clock_gettime(Process::CLOCK_MONOTONIC)
+  yield
+ensure
+  metric.increment(
+    by: Process.clock_gettime(Process::CLOCK_MONOTONIC) - start,
+    labels: labels,
+  )
+end
+
+n_string, n_concurrent_traces_string, = ARGV
+n, n_concurrent_traces = [n_string.to_i, n_concurrent_traces_string.to_i]
+
+n_concurrent_traces.times do
+  Prometheus::Client.tracer.send(:start, another_counter, {})
+end
+
+Benchmark.bm(7) do |x|
+  x.report("naive:") { n.times { trace(counter, worker: 1) { nil } } }
+  x.report("trace:") do
+    n.times { Prometheus::Client.tracer.trace(counter, worker: 1) { nil } }
+  end
+end
diff --git a/prometheus-client-tracer.gemspec b/prometheus-client-tracer.gemspec
index 3da9276..25a9c9f 100644
--- a/prometheus-client-tracer.gemspec
+++ b/prometheus-client-tracer.gemspec
@@ -23,4 +23,5 @@ Gem::Specification.new do |spec|
   spec.add_development_dependency "gc_ruboconfig", "= 2.4.0"
   spec.add_development_dependency "pry", "~> 0.10"
   spec.add_development_dependency "rspec", "~> 3.8"
+  spec.add_development_dependency "ruby-prof", "~> 0.18"
 end