__slots__ vs __dict__

[Jean Brouwers]

Classes using __slots__ seem to be quite a bit smaller and faster
to instantiate than regular Python classes using __dict__.

Below are the results for the __slots__ and __dict__ version of a
specific class with 16 attributes. Each line in the tables shows the
number of instances created so far, the total memory usage in Bytes,
the CPU time in secs, the average size per instance in Bytes and the
average CPU time per instance in micseconds.

Instances of this particular class with __slots__ are almost 6x
smaller and nearly 3x faster to create than intances of the __dict__
version. Results for other classes will vary, obviously.

Comments?

/Jean Brouwers
ProphICy Semiconductor, Inc.
PS) The tests were run on a dual 2.4 GHz Xeon system with RedHat
8.0 and Python 2.3.2. The test script is attached but keep in mind
that it only has been tested on Linux. It will not work elsewhere
due to the implementation of the memory() function.
testing __slots__ version ...
4096 insts so far: 3.0e+05 B 0.030 sec 73.0 B/i 7.3 usec/i
8192 insts so far: 8.8e+05 B 0.070 sec 107.5 B/i 8.5 usec/i
16384 insts so far: 1.5e+06 B 0.150 sec 92.2 B/i 9.2 usec/i
32768 insts so far: 3.3e+06 B 0.280 sec 101.0 B/i 8.5 usec/i
65536 insts so far: 6.6e+06 B 0.560 sec 101.2 B/i 8.5 usec/i
131072 insts so far: 1.4e+07 B 1.200 sec 103.4 B/i 9.2 usec/i
262144 insts so far: 2.7e+07 B 2.480 sec 103.4 B/i 9.5 usec/i
524288 insts so far: 5.5e+07 B 5.630 sec 104.0 B/i 10.7 usec/i
1048576 insts so far: 1.1e+08 B 13.980 sec 104.0 B/i 13.3 usec/i
1050000 insts total: 1.1e+08 B 14.000 sec 103.9 B/i 13.3 usec/i
testing __dict__ version ...
4096 insts so far: 2.4e+06 B 0.050 sec 595.0 B/i 12.2 usec/i
8192 insts so far: 4.6e+06 B 0.090 sec 564.5 B/i 11.0 usec/i
16384 insts so far: 9.5e+06 B 0.180 sec 581.8 B/i 11.0 usec/i
32768 insts so far: 1.9e+07 B 0.370 sec 582.2 B/i 11.3 usec/i
65536 insts so far: 3.8e+07 B 0.830 sec 582.6 B/i 12.7 usec/i
131072 insts so far: 7.6e+07 B 1.760 sec 582.7 B/i 13.4 usec/i
262144 insts so far: 1.5e+08 B 4.510 sec 582.8 B/i 17.2 usec/i
524288 insts so far: 3.1e+08 B 12.820 sec 582.8 B/i 24.5 usec/i
1048576 insts so far: 6.1e+08 B 38.370 sec 583.1 B/i 36.6 usec/i
1050000 insts total: 6.1e+08 B 38.380 sec 583.1 B/i 36.6 usec/i
-------------------------------slots.py-------------------------------
<pre>

from time import clock as time_clock
def cputime(since=0.0):
'''Return CPU in secs.
'''
return time_clock() - since
import os
_proc_status = '/proc/%d/status' % os.getpid() # Linux only
_scale = {'kB': 1024.0, 'mB': 1024.0*1024.0,
'KB': 1024.0, 'MB': 1024.0*1024.0}

def _VmB(VmKey):
global _scale
try: # get the /proc/<pid>/status pseudo file
t = open(_proc_status)
v = [v for v in t.readlines() if v.startswith(VmKey)]
t.close()
# convert Vm value to bytes
if len(v) == 1:
t = v[0].split() # e.g. 'VmRSS: 9999 kB'
if len(t) == 3: ## and t[0] == VmKey:
return float(t[1]) * _scale.get(t[2], 0.0)
except:
pass
return 0.0

def memory(since=0.0):
'''Return process memory usage in bytes.
'''
return _VmB('VmSize:') - since

def stacksize(since=0.0):
'''Return process stack size in bytes.
'''
return _VmB('VmStk:') - since

def slots(**kwds):
'''Return the slots names as sequence.
'''
return tuple(kwds.keys())

# __slots__ version
class SlotsClass(object):
__slots__ = slots(_attr1= False,
_attr2= None,
_attr3= None,
_attr4= None,
_attr5= None,
_attr6= None,
_attr7= 0,
_attr8= None,
_attr9= None,
_attr10=None,
_attr11=None,
_attr12=None,
_attr13=None,
_attr14=None,
_attr15=None,
_attr16=None)

def __init__(self, tuple4, parent):
self._attr1 = False
self._attr2 = None
self._attr3 = None
self._attr4 = None
self._attr5 = None
self._attr6 = None
if parent:
self._attr7 = parent._attr7 + 1
self._attr8 = parent._attr8
self._attr9 = parent._attr9
self._attr10 = parent
self._attr11 = parent._attr11
self._attr12 = parent._attr12
else:
self._attr7 = 0
self._attr8 = None
self._attr9 = None
self._attr10 = None
self._attr11 = self
self._attr12 = None
self._attr13, self._attr14, self._attr15, self._attr16 = tuple4
# __dict__ version
class DictClass(object):
_attr1 = None
_attr2 = None
_attr3 = None
_attr4 = None
_attr5 = None
_attr6 = None
_attr7 = 0
_attr8 = None
_attr9 = None
_attr10 = None
_attr11 = None
_attr12 = None
_attr13 = None
_attr14 = None
_attr15 = None
_attr16 = None

def __init__(self, tuple4, parent):
if parent:
self._attr7 = parent._attr7 + 1
self._attr8 = parent._attr8
self._attr9 = parent._attr9
self._attr10 = parent
self._attr11 = parent._attr11
self._attr12 = parent._attr12
else:
self._attr11 = self
self._attr13, self._attr14, self._attr15, self._attr16 = tuple4
if __name__ == '__main__':

import sys

def report(txt, n, b0, c0):
c = cputime(c0);
b = memory(b0)
print "%8d insts %s: %8.1e B %7.3f sec %6.1f B/i %6.1f usec/i" \
% (n, txt, b, c, b/n, 1.0e6*c/n)

if not sys.platform.startswith('linux'):
raise NotImplementedError, "%r not supported" % sys.platform

if 'dict' in sys.argv[1:]:
print 'testing __dict__ version ...'
testClass = DictClass
else:
print 'testing __slots__ version ...'
testClass = SlotsClass

t4 = ('', 0, 0, [])
b0 = memory()
c0 = cputime()
p = testClass(t4, None)
n, m = 1, 4096
# generate 1+ M instances
while n < 1050000: # 1048576:
p = testClass(t4, p)
n += 1
if n >= m: # occasionally print stats
m += m
report('so far', n, b0, c0)
report(' total', n, b0, c0)

</pre>

[Marcus von Appen]

Jean Brouwers <JB*******@ProphICy.com> writes:

Classes using __slots__ seem to be quite a bit smaller and faster
to instantiate than regular Python classes using __dict__.

[snip]

Yes, but instances usually will not have a Class.__dict__ attribute anymore.
Thus the following code throws an exception on binding new attributes on a
per-instance basis:

# start
class Foo (object):
__slots__ = "_test"

def __init__ (self):
self._test = None

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

Rebinding __slots__ in __setattr__ fails with a bus error on my system (it
should not be possible anyways, because __slots__ is a tuple):

# start
class Foo (object):
__slots__ = "_test", "_test2"

def __init__ (self):
self._test = 1
self._test2 = 2

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__ = self.__slots__
return

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

--- gdb backtrace

(gdb) r foo.py
....
Program received signal SIGBUS, Bus error.
0x08070050 in PyDict_GetItem ()
(gdb) bt
#0 0x08070050 in PyDict_GetItem ()
#1 0x08080ab7 in _PyType_Lookup ()
.....
Regards
Marcus

--
We don't understand the software, and sometimes we don't understand the
hardware, but we can *see* the blinking lights!

[Andrew Bennetts]

On Thu, May 13, 2004 at 12:07:54AM +0200, Marcus von Appen wrote:

Jean Brouwers <JB*******@ProphICy.com> writes:

Classes using __slots__ seem to be quite a bit smaller and faster
to instantiate than regular Python classes using __dict__. [snip]

Yes, but instances usually will not have a Class.__dict__ attribute anymore.
Thus the following code throws an exception on binding new attributes on a
per-instance basis:

# start
class Foo (object):
__slots__ = "_test"

def __init__ (self):
self._test = None

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

This code is a syntax error for me -- the final return isn't in a function.
Rebinding __slots__ in __setattr__ fails with a bus error on my system (it
should not be possible anyways, because __slots__ is a tuple):

# start
class Foo (object):
__slots__ = "_test", "_test2"

def __init__ (self):
self._test = 1
self._test2 = 2

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__ = self.__slots__
return

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

Also a syntax error. I also don't get a bus error, just infinite recursion
(because assigning to self.__slots__ calls __setattr__, but __slots__ isn't
an attribute of the instance). What version of Python, and what platform?
I'm guessing you're on something like FreeBSD where Python has had trouble
coping gracefully with infinite recursion.

-Andrew.

[Marcus von Appen]

Andrew Bennetts <an***************@puzzling.org> writes:

[...]

# start
class Foo (object):
__slots__ = "_test"

def __init__ (self):
self._test = None

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end
This code is a syntax error for me -- the final return isn't in a function.

Sorry for that messy code - the return is absolutely wrong (no idea, why I put
it there after typing the example in).

Rebinding __slots__ in __setattr__ fails with a bus error on my system (it
should not be possible anyways, because __slots__ is a tuple):

# start
class Foo (object):
__slots__ = "_test", "_test2"

def __init__ (self):
self._test = 1
self._test2 = 2

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__ = self.__slots__
return

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

Also a syntax error. I also don't get a bus error, just infinite recursion
(because assigning to self.__slots__ calls __setattr__, but __slots__ isn't
an attribute of the instance). What version of Python, and what platform?

Typed in a wrong line again here. Change it to something appropriate like

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__.__add__ (tuple (name))
return
I'm guessing you're on something like FreeBSD where Python has had trouble
coping gracefully with infinite recursion.

Right, but this was caused by typing in wrong code. Sorry for that.

Regards
Marcus

--
We don't understand the software, and sometimes we don't understand the
hardware, but we can *see* the blinking lights!

[Andrew Bennetts]

On Thu, May 13, 2004 at 06:55:39AM +0200, Marcus von Appen wrote:

Andrew Bennetts <an***************@puzzling.org> writes:

[...]

# start
class Foo (object):
__slots__ = "_test", "_test2"

def __init__ (self):
self._test = 1
self._test2 = 2

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__ = self.__slots__
return

if __name__ == "__main__":
f = Foo ()
f.testvar = "test"
return
# end

Also a syntax error. I also don't get a bus error, just infinite recursion
(because assigning to self.__slots__ calls __setattr__, but __slots__ isn't
an attribute of the instance). What version of Python, and what platform?

Typed in a wrong line again here. Change it to something appropriate like

def __setattr__ (self, name, value):
# just test a simple rebinding
self.__slots__.__add__ (tuple (name))
return

Now your example runs just fine (i.e. the script terminates normally, and
nothing happens). I still don't see any error. (And why do you keep
putting a redundant return at the end of your functions?)

Are you sure this is the same code that you get a crash with? What version
of Python are you using? (I've tested with 2.2.3 and 2.3.3).

-Andrew.

[Marcus von Appen]

Andrew Bennetts <an***************@puzzling.org> writes:

[...]

Now your example runs just fine (i.e. the script terminates normally, and
nothing happens). I still don't see any error.
I thought you would inspect the objects yourself. Let me show you another (now
working) example, which will explain it:

------
class DictClass (object):
test = "DictTest"

def __init__ (self):
pass

class SlotClass (object):
__slots__ = "test"

def __init__ (self):
self.test = "SlotTest"

def print_obj (obj):
# inspect .test and show the object attributes
print obj.test
print dir (obj)

# bind new attribute and inspect the object attributes
obj.new_test = "Test"
print obj.new_test
print dir (obj)
print "------"
return

if __name__ == "__main__":
dict_obj = DictClass ()
slot_obj = SlotClass ()

print_obj (dict_obj)
print_obj (slot_obj)

---

You will get something like the following output:

----------------
DictTest
['__class__', '__delattr__', '__dict__', ....., '__weakref__', 'test']

Test
['__class__', '__delattr__', '__dict__', ...., '__weakref__', 'new_test',
'test']

------

SlotTest
['__class__', '__delattr__', ...., '__slots__', '__str__', 'test']

Traceback (most recent call last):
File "foo.py", line 31, in ?
print_obj (slot_obj)
File "foo.py", line 19, in print_obj
obj.new_test = "Test"
AttributeError: 'SlotClass' object has no attribute 'new_test'
----------------

As you will notice, a __slot__ed object/class has no __weakref__ nor __dict__
attribute.
Thus binding new object attributes will fail with an AttributeError.

So let's try it again with using __setattr__ in SlotClass:

class SlotClass (object):
test = "SlotTest"
test2 = "SlotTest2" # just for making a tuple creation easier
__slots__ = test, test2

def __init__ (self):
pass

def __setattr__ (self, name, value):
# create a dict here to add its key to the __slot__ tuple
self.__slots__.__add__ (tuple ((name)))
return
def print_obj (obj):
# inspect .test and show the object attributes
print obj.test
print dir (obj)

# bind new attribute and inspect the object attributes
obj.new_test = "Test"
print obj.new_test
print dir (obj)
print "------"
return

if __name__ == "__main__":

#dict_obj = DictClass ()
slot_obj = SlotClass ()

#print_obj (dict_obj)
print_obj (slot_obj)

----
As you will see, you get an AttributeError again.tuple ((name))was not
concatenated to __slots__.
I think that should explain enough about the advantage and disadvantage of
__slots__ here.

(And why do you keep putting a redundant return at the end of your functions?)
I'm used to it :-).
Are you sure this is the same code that you get a crash with? What version
of Python are you using? (I've tested with 2.2.3 and 2.3.3).

No, I put messy code without _really_ thinking about it in my first post.
I did not realize that self.__slots__ = self._slots in __setattr__ will end
up in a recursion...

Regards
Marcus

--
We don't understand the software, and sometimes we don't understand the
hardware, but we can *see* the blinking lights!