The "popularity" of a memory word is the fraction of references to this word over the total of memory references in the execution of a program. In this paper we formally define the metric \emph{popularity of reference} and explain the reference patterns of the CommBench programs in terms of this metric. Our measurements indicate that embedded applications often exhibit a highly skewed distribution of popularity: a few thousand words are very popular while the rest of the data are referenced relatively few times. Based on the reference distribution profiles we discuss the allocation of data onto the top levels of the memory systems for embedded systems.