Introduce GEVENT_TRACK_GREENLET_TREE to disable greenlet tree features

As a performance optimization for applications where spawning
greenlets is critical. Plus some other optimizations to speed up
spawning in the general case.

CPython 3.6 with 1.2.2 vs these changes with tracking disabled:

| Benchmark              | 36_122_bench_spawn | 36config_bench_spawn_tree_off |
+------------------------+--------------------+-------------------------------+
| eventlet spawn         | 12.6 us            | 12.2 us: 1.04x faster (-4%)   |
| eventlet sleep         | 5.22 us            | 4.97 us: 1.05x faster (-5%)   |
| gevent spawn           | 4.27 us            | 5.06 us: 1.19x slower (+19%)  |
| gevent sleep           | 2.63 us            | 1.25 us: 2.11x faster (-53%)  |
| geventpool spawn       | 9.00 us            | 8.31 us: 1.08x faster (-8%)   |
| geventpool sleep       | 4.82 us            | 2.83 us: 1.70x faster (-41%)  |
| geventraw spawn        | 2.51 us            | 2.81 us: 1.12x slower (+12%)  |
| geventraw sleep        | 649 ns             | 679 ns: 1.05x slower (+5%)    |
| geventpool join        | 3.47 us            | 1.42 us: 2.44x faster (-59%)  |
| geventpool spawn kwarg | 11.0 us            | 8.95 us: 1.23x faster (-19%)  |
| geventraw spawn kwarg  | 3.87 us            | 4.20 us: 1.08x slower (+8%)   |

The differences compared to master are hard to quantify because the
standard deviation ends up being more than 10% of the mean in many
cases---and about a 10% improvement is what we typically see, so it
goes back and forth.

CHANGES.rst

@@ -14,6 +14,14 @@
   monkey-patch the underlying class, ``threading._Event``. Some code
   may be type-checking for that. See :issue:`1136`.
 
+- Introduce the configuration variable
+  `gevent.config.track_greenlet_tree` (aka
+  ``GEVENT_TRACK_GREENLET_TREE``) to allow disabling the greenlet tree
+  features for applications where greenlet spawning is performance
+  critical. This restores spawning performance to 1.2 levels.
+
+- Add additional optimizations for spawning greenlets, making it
+  faster than 1.3a2.
 
 1.3a2 (2018-03-06)
 ==================

src/gevent/_config.py

@@ -109,6 +109,8 @@ def validate_bool(value):
 def validate_anything(value):
     return value
 
+convert_str_value_as_is = validate_anything
+
 class Setting(object):
     name = None
     value = None
@@ -396,6 +398,27 @@ class TraceMalloc(Setting):
     tracking information for FFI objects.
     """
 
+
+class TrackGreenletTree(Setting):
+    name = 'track_greenlet_tree'
+    environment_key = 'GEVENT_TRACK_GREENLET_TREE'
+    default = True
+
+    validate = staticmethod(validate_bool)
+    # Don't do string-to-list conversion.
+    _convert = staticmethod(convert_str_value_as_is)
+
+    desc = """\
+    Should `Greenlet` objects track their spawning tree?
+
+    Setting this to a false value will make spawning `Greenlet`
+    objects and using `spawn_raw` faster, but the
+    ``spawning_greenlet``, ``spawn_tree_locals`` and ``spawning_stack``
+    will not be captured.
+
+    .. versionadded:: 1.3b1
+    """
+
 # The ares settings are all interpreted by
 # gevent/resolver/ares.pyx, so we don't do
 # any validation here.
@@ -410,7 +433,7 @@ class AresSettingMixin(object):
 
     validate = staticmethod(validate_anything)
 
-    _convert = staticmethod(validate_anything)
+    _convert = staticmethod(convert_str_value_as_is)
 
 class AresFlags(AresSettingMixin, Setting):
     name = 'ares_flags'

src/gevent/_greenlet.pxd

@@ -75,6 +75,9 @@ cdef inline list _extract_stack(int limit)
 @cython.locals(previous=_Frame, frame=tuple, f=_Frame)
 cdef _Frame _Frame_from_list(list frames)
 
+@cython.final
+cdef inline get_hub()
+
 
 cdef class Greenlet(greenlet):
     cdef readonly object value
@@ -128,7 +131,8 @@ cdef class Greenlet(greenlet):
 # doing a module global lookup. This is especially important
 # for spawning greenlets.
 cdef _greenlet__init__
-cdef get_hub
+cdef _threadlocal
+cdef get_hub_class
 cdef wref
 
 cdef Timeout
@@ -139,6 +143,7 @@ cdef wait
 cdef iwait
 cdef reraise
 cdef InvalidSwitchError
+cpdef GEVENT_CONFIG
 
 
 @cython.final

src/gevent/greenlet.py

@@ -16,10 +16,13 @@ from gevent._tblib import load_traceback
 from gevent.hub import GreenletExit
 from gevent.hub import InvalidSwitchError
 from gevent.hub import Waiter
-from gevent.hub import get_hub
+from gevent.hub import _threadlocal
+from gevent.hub import get_hub_class
 from gevent.hub import iwait
 from gevent.hub import wait
 from gevent.timeout import Timeout
+
+from gevent._config import config as GEVENT_CONFIG
 from gevent._util import Lazy
 from gevent._util import readproperty
 
@@ -39,6 +42,16 @@ __all__ = [
 locals()['getcurrent'] = __import__('greenlet').getcurrent
 locals()['greenlet_init'] = lambda: None
 
+def get_hub():
+    # This is identical to gevent.hub._get_hub_noargs so that it
+    # can be inlined for greenlet spawning by cython.
+    hub = _threadlocal.hub
+    if hub is None:
+        hubtype = get_hub_class()
+        hub = _threadlocal.hub = hubtype()
+    return hub
+
+
 if _PYPY:
     import _continuation # pylint:disable=import-error
     _continulet = _continuation.continulet
@@ -204,6 +217,12 @@ class Greenlet(greenlet):
             spawning greenlets, specify a larger value for improved debugging.
 
             .. versionadded:: 1.3a2
+
+        .. versionchanged:: 1.3b1
+           The ``GEVENT_TRACK_GREENLET_TREE`` configuration value may be set to
+           a false value to disable ``spawn_tree_locals``, ``spawning_greenlet``,
+           and ``spawning_stack``. The first two will be None in that case, and the
+           latter will be empty.
         """
         # greenlet.greenlet(run=None, parent=None)
         # Calling it with both positional arguments instead of a keyword
@@ -233,7 +252,6 @@ class Greenlet(greenlet):
 
 
         _greenlet__init__(self, None, get_hub())
-        #super(Greenlet, self).__init__(None, get_hub())
 
         if run is not None:
             self._run = run
@@ -266,6 +284,7 @@ class Greenlet(greenlet):
         # Failed with exception: (t, v, dump_traceback(tb)))
         self._exc_info = None
 
+        if GEVENT_CONFIG.track_greenlet_tree:
             spawner = getcurrent() # pylint:disable=undefined-variable
             self.spawning_greenlet = wref(spawner)
             try:
@@ -279,6 +298,12 @@ class Greenlet(greenlet):
 
             self._spawning_stack_frames = _extract_stack(self.spawning_stack_limit)
             self._spawning_stack_frames.extend(getattr(spawner, '_spawning_stack_frames', []))
+        else:
+            # None is the default for all of these in Cython, but we
+            # need to declare them for pure-Python mode.
+            self.spawning_greenlet = None
+            self.spawn_tree_locals = None
+            self._spawning_stack_frames = None
 
     @Lazy
     def spawning_stack(self):
@@ -286,7 +311,7 @@ class Greenlet(greenlet):
         # code. It's tempting to discard _spawning_stack_frames
         # after this, but child greenlets may still be created
         # that need it.
-        return _Frame_from_list(self._spawning_stack_frames)
+        return _Frame_from_list(self._spawning_stack_frames or [])
 
     def _get_minimal_ident(self):
         reg = self.parent.ident_registry

src/gevent/hub.py

@@ -27,7 +27,7 @@ __all__ = [
     'Waiter',
 ]
 
-from gevent._config import config
+from gevent._config import config as GEVENT_CONFIG
 from gevent._compat import string_types
 from gevent._compat import xrange
 from gevent._util import _NONE
@@ -63,8 +63,6 @@ get_ident = thread.get_ident
 MAIN_THREAD = get_ident()
 
 
-
-
 class LoopExit(Exception):
     """
     Exception thrown when the hub finishes running.
@@ -83,7 +81,6 @@ class LoopExit(Exception):
          affinity from a different thread
 
     """
-    pass
 
 
 class BlockingSwitchOutError(AssertionError):
@@ -101,6 +98,24 @@ class ConcurrentObjectUseError(AssertionError):
     # first one
     pass
 
+class TrackedRawGreenlet(RawGreenlet):
+
+    def __init__(self, function, parent):
+        RawGreenlet.__init__(self, function, parent)
+        # See greenlet.py's Greenlet class. We capture the cheap
+        # parts to maintain the tree structure, but we do not capture
+        # the stack because that's too expensive for 'spawn_raw'.
+
+        current = getcurrent()
+        self.spawning_greenlet = wref(current)
+        # See Greenlet for how trees are maintained.
+        try:
+            self.spawn_tree_locals = current.spawn_tree_locals
+        except AttributeError:
+            self.spawn_tree_locals = {}
+            if current.parent:
+                current.spawn_tree_locals = self.spawn_tree_locals
+
 
 def spawn_raw(function, *args, **kwargs):
     """
@@ -129,36 +144,31 @@ def spawn_raw(function, *args, **kwargs):
        Populate the ``spawning_greenlet`` and ``spawn_tree_locals``
        attributes of the returned greenlet.
 
+    .. versionchanged:: 1.3b1
+       *Only* populate ``spawning_greenlet`` and ``spawn_tree_locals``
+       if ``GEVENT_TRACK_GREENLET_TREE`` is enabled (the default). If not enabled,
+       those attributes will not be set.
+
     """
     if not callable(function):
         raise TypeError("function must be callable")
 
     # The hub is always the parent.
-    hub = get_hub()
+    hub = _get_hub_noargs()
+
+    factory = TrackedRawGreenlet if GEVENT_CONFIG.track_greenlet_tree else RawGreenlet
 
     # The callback class object that we use to run this doesn't
     # accept kwargs (and those objects are heavily used, as well as being
     # implemented twice in core.ppyx and corecffi.py) so do it with a partial
     if kwargs:
         function = _functools_partial(function, *args, **kwargs)
-        g = RawGreenlet(function, hub)
+        g = factory(function, hub)
         hub.loop.run_callback(g.switch)
     else:
-        g = RawGreenlet(function, hub)
+        g = factory(function, hub)
         hub.loop.run_callback(g.switch, *args)
 
-    # See greenlet.py's Greenlet class. We capture the cheap
-    # parts to maintain the tree structure, but we do not capture
-    # the stack because that's too expensive.
-    current = getcurrent()
-    g.spawning_greenlet = wref(current)
-    # See Greenlet for how trees are maintained.
-    try:
-        g.spawn_tree_locals = current.spawn_tree_locals
-    except AttributeError:
-        g.spawn_tree_locals = {}
-        if current.parent:
-            current.spawn_tree_locals = g.spawn_tree_locals
     return g
 
 
@@ -187,7 +197,7 @@ def sleep(seconds=0, ref=True):
 
     .. seealso:: :func:`idle`
     """
-    hub = get_hub()
+    hub = _get_hub_noargs()
     loop = hub.loop
     if seconds <= 0:
         waiter = Waiter()
@@ -209,7 +219,7 @@ def idle(priority=0):
 
     .. seealso:: :func:`sleep`
     """
-    hub = get_hub()
+    hub = _get_hub_noargs()
     watcher = hub.loop.idle()
     if priority:
         watcher.priority = priority
@@ -243,7 +253,7 @@ def kill(greenlet, exception=GreenletExit):
             # after it's been killed
             greenlet.kill(exception=exception, block=False)
         else:
-            get_hub().loop.run_callback(greenlet.throw, exception)
+            _get_hub_noargs().loop.run_callback(greenlet.throw, exception)
 
 
 class signal(object):
@@ -274,7 +284,7 @@ class signal(object):
         if not callable(handler):
             raise TypeError("signal handler must be callable.")
 
-        self.hub = get_hub()
+        self.hub = _get_hub_noargs()
         self.watcher = self.hub.loop.signal(signalnum, ref=False)
         self.watcher.start(self._start)
         self.handler = handler
@@ -393,6 +403,12 @@ def get_hub(*args, **kwargs):
 
     If a hub does not exist in the current thread, a new one is
     created of the type returned by :func:`get_hub_class`.
+
+    .. deprecated:: 1.3b1
+       The ``*args`` and ``**kwargs`` arguments are deprecated. They were
+       only used when the hub was created, and so were non-deterministic---to be
+       sure they were used, *all* callers had to pass them, or they were order-dependent.
+       Use ``set_hub`` instead.
     """
     hub = _threadlocal.hub
     if hub is None:
@@ -400,6 +416,15 @@ def get_hub(*args, **kwargs):
         hub = _threadlocal.hub = hubtype(*args, **kwargs)
     return hub
 
+def _get_hub_noargs():
+    # Just like get_hub, but cheaper to call because it
+    # takes no arguments or kwargs. See also a copy in
+    # gevent/greenlet.py
+    hub = _threadlocal.hub
+    if hub is None:
+        hubtype = get_hub_class()
+        hub = _threadlocal.hub = hubtype()
+    return hub
 
 def _get_hub():
     """Return the hub for the current thread.
@@ -428,7 +453,7 @@ def _config(default, envvar):
 
 hub_ident_registry = IdentRegistry()
 
-class Hub(RawGreenlet):
+class Hub(TrackedRawGreenlet):
     """
     A greenlet that runs the event loop.
 
@@ -457,7 +482,7 @@ class Hub(RawGreenlet):
 
 
     def __init__(self, loop=None, default=None):
-        RawGreenlet.__init__(self)
+        TrackedRawGreenlet.__init__(self, None, None)
         if hasattr(loop, 'run'):
             if default is not None:
                 raise TypeError("Unexpected argument: default")
@@ -476,7 +501,7 @@ class Hub(RawGreenlet):
             self.loop = self.loop_class(flags=loop, default=default) # pylint:disable=not-callable
         self._resolver = None
         self._threadpool = None
-        self.format_context = config.format_context
+        self.format_context = GEVENT_CONFIG.format_context
         self.minimal_ident = hub_ident_registry.get_ident(self)
 
     @Lazy
@@ -485,11 +510,11 @@ class Hub(RawGreenlet):
 
     @property
     def loop_class(self):
-        return config.loop
+        return GEVENT_CONFIG.loop
 
     @property
     def backend(self):
-        return config.libev_backend
+        return GEVENT_CONFIG.libev_backend
 
     def __repr__(self):
         if self.loop is None:
@@ -741,7 +766,7 @@ class Hub(RawGreenlet):
 
     @property
     def resolver_class(self):
-        return config.resolver
+        return GEVENT_CONFIG.resolver
 
     def _get_resolver(self):
         if self._resolver is None:
@@ -759,7 +784,7 @@ class Hub(RawGreenlet):
 
     @property
     def threadpool_class(self):
-        return config.threadpool
+        return GEVENT_CONFIG.threadpool
 
     def _get_threadpool(self):
         if self._threadpool is None:
@@ -820,7 +845,7 @@ class Waiter(object):
     __slots__ = ['hub', 'greenlet', 'value', '_exception']
 
     def __init__(self, hub=None):
-        self.hub = get_hub() if hub is None else hub
+        self.hub = _get_hub_noargs() if hub is None else hub
         self.greenlet = None
         self.value = None
         self._exception = _NONE
@@ -963,7 +988,7 @@ def iwait(objects, timeout=None, count=None):
     """
     # QQQ would be nice to support iterable here that can be generated slowly (why?)
     if objects is None:
-        yield get_hub().join(timeout=timeout)
+        yield _get_hub_noargs().join(timeout=timeout)
         return
 
     count = len(objects) if count is None else min(count, len(objects))
@@ -971,7 +996,7 @@ def iwait(objects, timeout=None, count=None):
     switch = waiter.switch
 
     if timeout is not None:
-        timer = get_hub().loop.timer(timeout, priority=-1)
+        timer = _get_hub_noargs().loop.timer(timeout, priority=-1)
         timer.start(switch, _NONE)
 
     try:
@@ -1031,7 +1056,7 @@ def wait(objects=None, timeout=None, count=None):
     .. seealso:: :func:`iwait`
     """
     if objects is None:
-        return get_hub().join(timeout=timeout)
+        return _get_hub_noargs().join(timeout=timeout)
     return list(iwait(objects, timeout, count))
 
 

src/gevent/util.py

@@ -283,6 +283,10 @@ class GreenletTree(object):
         tb = ''.join(traceback.format_stack(frame))
         tree.child_multidata(tb)
 
+    @staticmethod
+    def __spawning_parent(greenlet):
+        return (getattr(greenlet, 'spawning_greenlet', None) or _noop)()
+
     def __render_locals(self, tree):
         gr_locals = all_local_dicts_for_greenlet(self.greenlet)
         if gr_locals:
@@ -316,7 +320,7 @@ class GreenletTree(object):
         if spawning_stack and tree.details and tree.details['stacks']:
             self.__render_tb(tree, 'Spawned at:', spawning_stack)
 
-        spawning_parent = getattr(self.greenlet, 'spawning_greenlet', _noop)()
+        spawning_parent = self.__spawning_parent(self.greenlet)
         tree_locals = getattr(self.greenlet, 'spawn_tree_locals', None)
         if tree_locals and tree_locals is not getattr(spawning_parent, 'spawn_tree_locals', None):
             tree.child_data('Spawn Tree Locals')
@@ -377,7 +381,7 @@ class GreenletTree(object):
             if not isinstance(ob, RawGreenlet):
                 continue
 
-            spawn_parent = getattr(ob, 'spawning_greenlet', _noop)()
+            spawn_parent = cls.__spawning_parent(ob)
 
             if spawn_parent is None:
                 root = cls._root_greenlet(ob)

src/greentest/test__util.py

@@ -59,15 +59,22 @@ class TestFormat(greentest.TestCase):
 @greentest.skipOnPyPy("See TestFormat")
 class TestTree(greentest.TestCase):
 
-    @greentest.ignores_leakcheck
-    def test_tree(self):
+    def setUp(self):
+        super(TestTree, self).setUp()
+        self.track_greenlet_tree = gevent.config.track_greenlet_tree
+        gevent.config.track_greenlet_tree = True
+
+    def tearDown(self):
+        gevent.config.track_greenlet_tree = self.track_greenlet_tree
+        super(TestTree, self).tearDown()
+
+    def _build_tree(self):
         # pylint:disable=too-many-locals
         # Python 2.7 on Travis seems to show unexpected greenlet objects
         # so perhaps we need a GC?
-        gc.collect()
+        for _ in range(3):
             gc.collect()
 
-        import re
         glets = []
         l = MyLocal(42)
         assert l
@@ -96,6 +103,8 @@ class TestTree(greentest.TestCase):
             return s(t2)
 
         s3 = s(t3)
+        if s3.spawn_tree_locals is not None:
+            # Can only do this if we're tracking spawn trees
             s3.spawn_tree_locals['stl'] = 'STL'
         s3.join()
 
@@ -104,10 +113,17 @@ class TestTree(greentest.TestCase):
         s4.join()
 
         tree = s4.value
+        return tree, str(tree), tree.format(details={'stacks': False})
+
+    @greentest.ignores_leakcheck
+    def test_tree(self):
+        import re
+        tree, str_tree, tree_format = self._build_tree()
+
         self.assertTrue(tree.root)
 
-        self.assertNotIn('Parent', str(tree)) # Simple output
-        value = tree.format(details={'stacks': False})
+        self.assertNotIn('Parent', str_tree) # Simple output
+        value = tree_format
         hexobj = re.compile('0x[0123456789abcdef]+L?', re.I)
         value = hexobj.sub('X', value)
         value = value.replace('epoll', 'select')
@@ -145,5 +161,11 @@ class TestTree(greentest.TestCase):
         """.strip()
         self.assertEqual(value, expected)
 
+    @greentest.ignores_leakcheck
+    def test_tree_no_track(self):
+        gevent.config.track_greenlet_tree = False
+        self._build_tree()
+
+
 if __name__ == '__main__':
     greentest.main()