diff --git a/clcache.py b/clcache.py
index 860c28dd..5fcc8219 100644
--- a/clcache.py
+++ b/clcache.py
@@ -17,6 +17,7 @@
 import hashlib
 import json
 import multiprocessing
+import msvcrt
 import os
 import pickle
 import re
@@ -31,6 +32,7 @@
 HashAlgorithm = hashlib.md5
 
 OUTPUT_LOCK = threading.Lock()
+PROCESS_LOCK = threading.Lock()
 
 # try to use os.scandir or scandir.scandir
 # fall back to os.listdir if not found
@@ -67,6 +69,7 @@
 # Define some Win32 API constants here to avoid dependency on win32pipe
 NMPWAIT_WAIT_FOREVER = wintypes.DWORD(0xFFFFFFFF)
 ERROR_PIPE_BUSY = 231
+HANDLE_FLAG_INHERIT = 1
 
 # ManifestEntry: an entry in a manifest file
 # `includeFiles`: list of paths to include files, which this source file uses
@@ -1324,13 +1327,26 @@ def invokeRealCompiler(compilerBinary, cmdLine, captureOutput=False, outputAsStr
     if captureOutput:
         # Don't use subprocess.communicate() here, it's slow due to internal
         # threading.
-        with TemporaryFile() as stdoutFile, TemporaryFile() as stderrFile:
+        with PROCESS_LOCK:
+            stdoutFile = TemporaryFile()
+            stderrFile = TemporaryFile()
+
+            # Make temporary file handles non-inheritable, this is a workaround to
+            # a problem in Python 3.3
+            windll.kernel32.SetHandleInformation(msvcrt.get_osfhandle(stdoutFile.fileno()), HANDLE_FLAG_INHERIT, 0)
+            windll.kernel32.SetHandleInformation(msvcrt.get_osfhandle(stderrFile.fileno()), HANDLE_FLAG_INHERIT, 0)
+
             compilerProcess = subprocess.Popen(realCmdline, stdout=stdoutFile, stderr=stderrFile, env=environment)
+
+        try:
             returnCode = compilerProcess.wait()
             stdoutFile.seek(0)
             stdout = stdoutFile.read()
             stderrFile.seek(0)
             stderr = stderrFile.read()
+        finally:
+            stdoutFile.close()
+            stderrFile.close()
     else:
         returnCode = subprocess.call(realCmdline, env=environment)
 
diff --git a/integrationtests.py b/integrationtests.py
index bcef1d45..eab2538a 100644
--- a/integrationtests.py
+++ b/integrationtests.py
@@ -807,18 +807,19 @@ def testHitsViaMpConcurrent(self):
 
     def testOutput(self):
         # type: () -> None
-        with cd(os.path.join(ASSETS_DIR, "parallel")), tempfile.TemporaryDirectory() as tempDir:
-            sources = glob.glob("*.cpp")
-            clcache.Cache(tempDir)
-            customEnv = self._createEnv(tempDir)
-            cmd = CLCACHE_CMD + ["/nologo", "/EHsc", "/c"]
-            mpFlag = "/MP" + str(len(sources))
-            out = subprocess.check_output(cmd + [mpFlag] + sources, env=customEnv).decode("ascii")
-            # print the output so that it shows up in py.test
-            print(out)
-
-            for s in sources:
-                self.assertEqual(out.count(s), 1)
+        for _ in range(100):
+            with cd(os.path.join(ASSETS_DIR, "parallel")), tempfile.TemporaryDirectory() as tempDir:
+                sources = glob.glob("*.cpp")
+                clcache.Cache(tempDir)
+                customEnv = self._createEnv(tempDir)
+                cmd = CLCACHE_CMD + ["/nologo", "/EHsc", "/c"]
+                mpFlag = "/MP" + str(len(sources))
+                out = subprocess.check_output(cmd + [mpFlag] + sources, env=customEnv).decode("ascii")
+                # print the output so that it shows up in py.test
+                print(out)
+
+                for s in sources:
+                    self.assertEqual(out.count(s), 1)
 
 class TestRunParallel(RunParallelBase, unittest.TestCase):
     env = dict(os.environ)