A quantitative analysis of locality in dataflow programs

Abstract

Substantial evidence suggests that exploiting some forms of locality within datajiow programs can impact performance dramatically. This is the basic premise of several hybrid von Neumann-dataflow or multithreaded architectures. Identifying and exploiting locality, however, in a jine-grained asynchronous execution model is not trivial. In this paper, jine grained intra-thread locality is defined, quantified and evaitiated. These experimental measurements are based on the evaluation of a set of numer+c and non-numeric benchmarks. The results point to a very large degree of thread locality: for example, over 70% of the instructions have to wait tess than 5 instruction execution steps for their input data. Furthermore, the remarkable uniformity and consistency of the distti”bution of thread locality across a wide vam”ety of benchmarks suggests that thread locality is highly dependent on the instruction set.