There likely is an optimal stride, as a function of the platform and the data you are storing. There is also a worst-case stride. You probably don't know what the good and bad strides are, and because they are data-dependent, they can change over time. Randomness gives you a very high probability of avoiding very bad case performance. And (to rephrase what another comment says) avoiding a constant stride avoids unlucky 'resonances' where the interval interacts with some other fixed property of your system or data.