OK, I was struck my the significant difference Falkkin obtained. I ran the same script on my development box here (a Sun Ultra 2, with a 300MHz Ultrasparc CPU and 128 MB of ram), and these were my results:

odd_1:  9 wallclock secs ( 8.55 usr +  0.00 sys =  8.55 CPU)
odd_2:  9 wallclock secs ( 9.46 usr +  0.00 sys =  9.46 CPU)
[download]

The divide is about 10% faster than the shift. Why is this so? I personally would have thought a shift operation - a basic opcode in just about any instruction set - would be faster that a division.

Update: I followed runrig's advice and took out the rand call. The results were different:

odd_1:  7 wallclock secs ( 6.41 usr +  0.00 sys =  6.41 CPU)
odd_2:  7 wallclock secs ( 6.00 usr +  0.00 sys =  6.00 CPU)
[download]

Now the division is longer - by about 7%. Strange stuff.