It depends on how/why you are measuring transaction speed. If it is for publication, you probably need to have an independant lab do it (your own numbers would be meaningless). If you are just trying to get a yardstick on performance, you can probably just time some transactions and compute the average of them.
The first loop isolates the amount of time used by the loop itself, which should be relatively small. The second loop repeats the code you wish to benchmark. When you are done, you subtract the overhead from the test time and divide by the number of iterations to compute the average.