How much better are python local variables over globals, attributes, or slots?

David A. Kra — Fri, 27 Dec 2024 09:57:34 +0000

Alternatively, how much worse are the others compared to local variables?

What's the big deal?

Some articles advise python programmers to use local variables if they will be used a lot, even if only referenced but not changed.

For example, here are two quotes from "Making Python Programs Blazing Fast" :

"There's actually difference in speed of lookup even between - let's say - local variable in function (fastest), class-level attribute (e.g. self.name - slower) and global for example imported function like time.time (slowest)."

"Don't Access Attributes

Another thing that might slow down your programs is dot operator (.) which is used when accessing object attributes. This operator triggers dictionary lookup using getattribute, which creates extra overhead in your code. "

Also, these are not the only possibilities. What about differences among:

Class attributes,
Instance attributes,
Slot Instance attributes,
Slot Instance attributes of a different instance
Method attributes,
Function (non-method) attributes,
And as compared to using a constant?

Testing

I tested by using all these in loops with 1,000,000 iterations, but minimal work per iteration inside them. The functions were all methods of a class, except for the function used to test function attributes for functions outside a class.

A typical function was of the form:

      def testINSA(self):
          s=0
          for i in range(10000000):
            s+=self.INSA # what is after the = differs in each function.
          return s

All the code is at the bottom of this post.

Results

Addend location	Avg time of 10 calls	Rate per Second	Relative Performance
LV Local Variable	1.0253	0.975	100%
C Constant	1.0896	0.918	94%
ISA Instance Slotted Attribute	1.1156	0.896	92%
GV Global Variable	1.1293	0.886	91%
OISA Other Instance Slotted Attribute	1.2055	0.830	85%
CA Class Attribute	1.6385	0.610	63%
INSA Instance Non-Slotted Attribute	1.739	0.575	59%
FA Function Attribute	1.8076	0.553	57%
MA Method Attribute	3.3259	0.301	31%

Explanation:
When comparing performance, always compare the (Achievements / Resource), such as Miles per Gallon, or Calculations per Second, rather than Liters per Km, or Seconds per Calculation.

Time is as reported by timeit, limited to 3 significant digits.
Rate is the reciprocal of time. How many of these per second.
Relative performance is the rate as compared to the best. By using relative performance, bigger is better, and twice as big is twice as good.

Findings

Yes, the local variables are the fastest. They are even faster than constants.

The performance of the different types of variables is clustered into groupings.

Locals do perform the best.
Constants, class instance slotted attributes, and global variables: Constants were 94% as fast, Class Instance slotted attributes were 92% as fast, while Global variables were 91% as fast as locals.
Other Instance slotted attributes: These are slotted instance attributes of some other object. These performed 85% of a local variable.
Class Attribute, Class Instance Non-Slotted Attribute, and Function Attribute: These performed, respectively, at 63%, 59%, and 57% of a local variable.
Method attribute A method attribute performs in a class of its own, at 31% of the performance of a local variable.

Conclusions

If used millions of times, it is worthwhile to use local variables and slot attributes for performance reasons, not merely for code quality reasons.

The surprise is that compared to common wisdom, globals are almost as performant as slotted instance attributes in a class. Another surprise is that method attributes are the worst.

Code