If I was doing this, I would work from the baud rate and the message sizes to estimate how long it takes to do one "cycle" of the algorithm--that is, how long it takes to transmit the mask and the test address and then to receive the slaves' responses. Then I would modify the simulator program that I posted to count and display the number of cycles used. Then I would run the program with various combinations of slave addresses to see how it performed under different conditions.

As Erik has mentioned, the worst situation is when the slave addresses differ only in their least significant bits. When that's true, the algorithm has to search all the way to the bottom of the tree before it can distinguish the slaves. So to get worst case timing, you would want to specify some number of consecutive slave addresses. To get more "typical" timing, you might specify some number of random slave addresses.

Please note that the algorithm ASSUMES that you can reliably distinguish between "no response", "one response", and "multiple responses". As others have pointed out, this may not be as easy as it sounds. Your first task should probably be to make sure you can do that. It would be silly to waste time estimating the performance of an algorithm that you can't use.

-- Russ