looking for a loop optimization example

I'm working on a PyMOTW post about the dis module. As one of the examples, I'd like to show how to use dis to reduce the number of opcodes processed in a loop as an optimization technique. I'm looking for some "naive" code that looks OK, but doesn't perform especially well on large datasets.

So far I have a class that reads /usr/share/dict/words and organizes the words by their first letter:

#!/usr/bin/env python
# encoding: utf-8

import timeit
import dis

class Dictionary(object):
    def __init__(self, filename='/usr/share/dict/words'):
        self.by_letter = {}
        self.load_file(filename)
    def load_file(self, filename):
        f = open(filename, 'rt')
        try:
            words = [l.strip() for l in f.readlines()]
            for word in words:
                try:
                    self.by_letter[word[0]].append(word)
                except KeyError:
                    self.by_letter[word[0]] = [word]
        finally:
            f.close()
        
if __name__ == '__main__':
    print 'DISASSEMBLY:'
    dis.dis(Dictionary)
    print
    t = timeit.Timer('d = Dictionary()', 'from dis_slow_loop import Dictionary')
    iterations = 2
    print 'TIME: %0.4f' % (t.timeit(iterations)/iterations)

I've optimized it a bit to get this:

#!/usr/bin/env python
# encoding: utf-8

import collections
import dis
import timeit

class Dictionary(object):
    def __init__(self, filename='/usr/share/dict/words'):
        self.by_letter = collections.defaultdict(list)
        self.load_file(filename)
    def load_file(self, filename):
        by_letter = self.by_letter
        with open(filename, 'rt') as f:
            for line in f:
                word = line.strip()
                by_letter[word[0]].append(word)

if __name__ == '__main__':
    print 'DISASSEMBLY:'
    dis.dis(Dictionary)
    print
    t = timeit.Timer('d = Dictionary()', 'from dis_faster_loop import Dictionary')
    iterations = 2
    print 'TIME: %0.4f' % (t.timeit(iterations)/iterations)

The first example gives me a run time of 0.2712 sec and the second is 0.1998.

I'm looking for either alternative loops or a way to increase that spread without making the "slow" version obviously obtuse (no sleeps, unnecessary sub-loops, etc.).