diff --git a/tools/run-deopt-fuzzer.py b/tools/run-deopt-fuzzer.py
index 2b9fe1e254..b3ea2aa42d 100755
--- a/tools/run-deopt-fuzzer.py
+++ b/tools/run-deopt-fuzzer.py
@@ -373,7 +373,8 @@ def Execute(arch, mode, args, options, suites, workspace):
                         options.random_seed,
                         True,  # No sorting of test cases.
                         0,  # Don't rerun failing tests.
-                        0)  # No use of a rerun-failing-tests maximum.
+                        0,  # No use of a rerun-failing-tests maximum.
+                        False)  # No predictable mode.
 
   # Find available test suites and read test cases from them.
   variables = {
diff --git a/tools/run-tests.py b/tools/run-tests.py
index 03ad716c67..dc09357957 100755
--- a/tools/run-tests.py
+++ b/tools/run-tests.py
@@ -171,6 +171,9 @@ def BuildOptions():
                     help="Comma-separated list of testing variants")
   result.add_option("--outdir", help="Base directory with compile output",
                     default="out")
+  result.add_option("--predictable",
+                    help="Compare output of several reruns of each test",
+                    default=False, action="store_true")
   result.add_option("-p", "--progress",
                     help=("The style of progress indicator"
                           " (verbose, dots, color, mono)"),
@@ -301,6 +304,11 @@ def ProcessOptions(options):
     options.flaky_tests = "skip"
     options.slow_tests = "skip"
     options.pass_fail_tests = "skip"
+  if options.predictable:
+    VARIANTS = ["default"]
+    options.extra_flags.append("--predictable")
+    options.extra_flags.append("--verify_predictable")
+    options.extra_flags.append("--no-inline-new")
 
   if not options.shell_dir:
     if options.shell:
@@ -416,6 +424,11 @@ def Execute(arch, mode, args, options, suites, workspace):
       timeout = TIMEOUT_DEFAULT;
 
   timeout *= TIMEOUT_SCALEFACTOR[mode]
+
+  if options.predictable:
+    # Predictable mode is slower.
+    timeout *= 2
+
   ctx = context.Context(arch, mode, shell_dir,
                         mode_flags, options.verbose,
                         timeout, options.isolates,
@@ -425,7 +438,8 @@ def Execute(arch, mode, args, options, suites, workspace):
                         options.random_seed,
                         options.no_sorting,
                         options.rerun_failures_count,
-                        options.rerun_failures_max)
+                        options.rerun_failures_max,
+                        options.predictable)
 
   # TODO(all): Combine "simulator" and "simulator_run".
   simulator_run = not options.dont_skip_simulator_slow_tests and \
diff --git a/tools/testrunner/local/execution.py b/tools/testrunner/local/execution.py
index 43d11ebe8a..d2d086c111 100644
--- a/tools/testrunner/local/execution.py
+++ b/tools/testrunner/local/execution.py
@@ -126,6 +126,63 @@ class Runner(object):
       pool.add([self._GetJob(test)])
       self.remaining += 1
 
+  def _ProcessTestNormal(self, test, result, pool):
+    self.indicator.AboutToRun(test)
+    test.output = result[1]
+    test.duration = result[2]
+    has_unexpected_output = test.suite.HasUnexpectedOutput(test)
+    if has_unexpected_output:
+      self.failed.append(test)
+      if test.output.HasCrashed():
+        self.crashed += 1
+    else:
+      self.succeeded += 1
+    self.remaining -= 1
+    self.indicator.HasRun(test, has_unexpected_output)
+    if has_unexpected_output:
+      # Rerun test failures after the indicator has processed the results.
+      self._MaybeRerun(pool, test)
+    # Update the perf database if the test succeeded.
+    return not has_unexpected_output
+
+  def _ProcessTestPredictable(self, test, result, pool):
+    # Always pass the test duration for the database update.
+    test.duration = result[2]
+    if test.run == 1 and result[1].HasTimedOut():
+      # If we get a timeout in the first run, we are already in an
+      # unpredictable state. Just report it as a failure and don't rerun.
+      self.indicator.AboutToRun(test)
+      test.output = result[1]
+      self.remaining -= 1
+      self.failed.append(test)
+      self.indicator.HasRun(test, True)
+    if test.run > 1 and test.output != result[1]:
+      # From the second run on, check for differences. If a difference is
+      # found, call the indicator twice to report both tests. All runs of each
+      # test are counted as one for the statistic.
+      self.indicator.AboutToRun(test)
+      self.remaining -= 1
+      self.failed.append(test)
+      self.indicator.HasRun(test, True)
+      self.indicator.AboutToRun(test)
+      test.output = result[1]
+      self.indicator.HasRun(test, True)
+    elif test.run >= 3:
+      # No difference on the third run -> report a success.
+      self.indicator.AboutToRun(test)
+      self.remaining -= 1
+      self.succeeded += 1
+      test.output = result[1]
+      self.indicator.HasRun(test, False)
+    else:
+      # No difference yet and less than three runs -> add another run and
+      # remember the output for comparison.
+      test.run += 1
+      test.output = result[1]
+      pool.add([self._GetJob(test)])
+    # Always update the perf database.
+    return True
+
   def Run(self, jobs):
     self.indicator.Starting()
     self._RunInternal(jobs)
@@ -156,22 +213,12 @@ class Runner(object):
       it = pool.imap_unordered(RunTest, queue)
       for result in it:
         test = test_map[result[0]]
-        self.indicator.AboutToRun(test)
-        test.output = result[1]
-        test.duration = result[2]
-        has_unexpected_output = test.suite.HasUnexpectedOutput(test)
-        if has_unexpected_output:
-          self.failed.append(test)
-          if test.output.HasCrashed():
-            self.crashed += 1
+        if self.context.predictable:
+          update_perf = self._ProcessTestPredictable(test, result, pool)
         else:
+          update_perf = self._ProcessTestNormal(test, result, pool)
+        if update_perf:
           self._RunPerfSafe(lambda: self.perfdata.UpdatePerfData(test))
-          self.succeeded += 1
-        self.remaining -= 1
-        self.indicator.HasRun(test, has_unexpected_output)
-        if has_unexpected_output:
-          # Rerun test failures after the indicator has processed the results.
-          self._MaybeRerun(pool, test)
     finally:
       pool.terminate()
       self._RunPerfSafe(lambda: self.perf_data_manager.close())
diff --git a/tools/testrunner/objects/context.py b/tools/testrunner/objects/context.py
index d288e0e053..937d9089f3 100644
--- a/tools/testrunner/objects/context.py
+++ b/tools/testrunner/objects/context.py
@@ -29,7 +29,8 @@
 class Context():
   def __init__(self, arch, mode, shell_dir, mode_flags, verbose, timeout,
                isolates, command_prefix, extra_flags, noi18n, random_seed,
-               no_sorting, rerun_failures_count, rerun_failures_max):
+               no_sorting, rerun_failures_count, rerun_failures_max,
+               predictable):
     self.arch = arch
     self.mode = mode
     self.shell_dir = shell_dir
@@ -44,16 +45,17 @@ class Context():
     self.no_sorting = no_sorting
     self.rerun_failures_count = rerun_failures_count
     self.rerun_failures_max = rerun_failures_max
+    self.predictable = predictable
 
   def Pack(self):
     return [self.arch, self.mode, self.mode_flags, self.timeout, self.isolates,
             self.command_prefix, self.extra_flags, self.noi18n,
             self.random_seed, self.no_sorting, self.rerun_failures_count,
-            self.rerun_failures_max]
+            self.rerun_failures_max, self.predictable]
 
   @staticmethod
   def Unpack(packed):
     # For the order of the fields, refer to Pack() above.
     return Context(packed[0], packed[1], None, packed[2], False,
                    packed[3], packed[4], packed[5], packed[6], packed[7],
-                   packed[8], packed[9], packed[10], packed[11])
+                   packed[8], packed[9], packed[10], packed[11], packed[12])
diff --git a/tools/testrunner/objects/output.py b/tools/testrunner/objects/output.py
index 87b4c84e19..6b5b0cd491 100644
--- a/tools/testrunner/objects/output.py
+++ b/tools/testrunner/objects/output.py
@@ -38,6 +38,12 @@ class Output(object):
     self.stdout = stdout
     self.stderr = stderr
 
+  def __ne__(self, other):
+    return (self.exit_code != other.exit_code or
+            self.timed_out != other.timed_out or
+            self.stdout != other.stdout or
+            self.stderr != other.stderr)
+
   def HasCrashed(self):
     if utils.IsWindows():
       return 0x80000000 & self.exit_code and not (0x3FFFFF00 & self.exit_code)