Add option to enable cProfile

charm4py/charm.py

@@ -28,6 +28,7 @@
 from . import reduction
 from . import wait
 import array
+import cProfile
 try:
     import numpy
 except ImportError:
@@ -96,6 +97,7 @@ def __init__(self):
 
         self.options = Options()
         self.options.profiling = False
+        self.options.cprofiling = False
         self.options.pickle_protocol = -1  # -1 selects the highest protocol number
         self.options.local_msg_optim = True
         self.options.local_msg_buf_size = 50
@@ -292,7 +294,8 @@ def recvArrayMsg(self, aid, index, ep, msg, dcopy_start):
             if isinstance(args, Chare):  # obj migrating in
                 em = self.entryMethods[ep + 1]  # get 'migrated' EntryMethod object instead of __init__
                 obj = args
-                obj._contributeInfo = self.lib.initContributeInfo(aid, index, CONTRIBUTOR_TYPE_ARRAY)
+                obj._contributeInfo = self.lib.initContributeInfo(
+                    aid, index, CONTRIBUTOR_TYPE_ARRAY)
                 self.arrays[aid][index] = obj
                 em.run(obj, {}, ())
             else:
@@ -992,7 +995,10 @@ def getPeHost(self, pe):
     def getPeHostRank(self, pe):
         return self.lib.CkPhysicalRank(pe)
 
-    def exit(self, exit_code=0):
+    def exit(self, exit_code=0, profile_filename="profile"):
+        if self.options.cprofiling:
+            self.thisProxy.stop_cprofile(filename=profile_filename,
+                                         awaitable=True).get()
         self.lib.CkExit(exit_code)
 
     def abort(self, message):
@@ -1012,6 +1018,8 @@ class CharmRemote(Chare):
 
     def __init__(self):
         charm.thisProxy = self.thisProxy
+        if charm.options.cprofiling:
+            self.start_cprofile()
 
     def exit(self, exit_code=0):
         charm.exit(exit_code)
@@ -1028,6 +1036,16 @@ def LBTurnInstrumentOff(self):
     def addReducer(self, func):
         charm.addReducer(func)
 
+    def start_cprofile(self):
+        self.pr = cProfile.Profile()
+        self.pr.enable()
+
+    def stop_cprofile(self, dump=True, filename="profile"):
+        if hasattr(self, "pr"):
+            self.pr.disable()
+            if dump:
+                self.pr.dump_stats(filename + str(self.myPe()) + ".prof")
+
     # user signature is: `def updateGlobals(self, global_dict, module_name='__main__')`
     def updateGlobals(self, *args):
         global_dict = {}

docs/profiling.rst

@@ -31,7 +31,7 @@ lot of time waiting for communication to complete (which depends on the size
 and number of messages per second, network throughput and latency, efficiency
 of the communication layer, etc.). If you are seeing significant idle time, one
 thing that can increase performance is using an efficient communication layer
-such as MPI (see :doc:`perf-tips`).
+such as MPI (see :doc:`perf-tips`). Charm4Py also supports profiling using cProfile.
 
 In the future, we plan to support Projections_ which is a full-fledged
 parallel performance analysis tool for Charm++.
@@ -120,6 +120,10 @@ method of each Cell. Most of the send overhead is from the "run" method
 is called when one of its channels receives a message). There is also some
 receive overhead due to chares migrating into this PE (method ``migrated``).
 
-
+To use cProfile, the extension modules need to be rebuilt with profiling enabled. 
+This can be done by setting the ``CHARM4PY_BUILD_CPROFILE`` environment variable.
+To enable cProfile, set ``charm.options.cprofiling`` to ``True`` before the runtime
+starts. This will dump a profiling output file for each PE.
+Note that the entry method that calls ``charm.exit()`` needs to be made a coroutine.
 
 .. _Projections: https://charm.readthedocs.io/en/latest/projections/manual.html

setup.py

@@ -279,14 +279,20 @@ def run(self):
             else:
                 extra_link_args=["-Wl,-rpath,$ORIGIN/../.libs"]
 
+        compiler_directives = {}
+        if os.environ.get('CHARM4PY_BUILD_CPROFILE') == '1':
+            compiler_directives = {'profile': True}
+
         extensions.extend(cythonize(setuptools.Extension('charm4py.charmlib.charmlib_cython',
                               sources=['charm4py/charmlib/charmlib_cython.pyx'],
                               include_dirs=['charm_src/charm/include'] + my_include_dirs,
                               library_dirs=[os.path.join(os.getcwd(), 'charm4py', '.libs')],
                               libraries=["charm"],
                               extra_compile_args=['-g0', '-O3'],
                               extra_link_args=extra_link_args,
-                              ), compile_time_env={'HAVE_NUMPY': haveNumpy}))
+                              ),
+                            compiler_directives=compiler_directives,
+                            compile_time_env={'HAVE_NUMPY': haveNumpy}))
     else:
         try:
             check_cffi()