Aggregated Dynamic Dataflow Graph Generation and Visualization

Abstract—Aggregated Dynamic Dataflow Graphs can assist programmers to uncover the main data paths of a given algorithm. This information can be useful when scaling a singlethreaded program into a multi-core architecture. The amount of data movements is crucial when targeting for cache incoherent and/or heterogeneous platforms. This paper presents two methods for generating function-level Aggregated Dynamic Dataflow Graphs. Instruction level trace log was used as a basis, which was generated by Microsoft Giano processor simulator platform. Top-down aggregation strategy and relational database was used to speed up the generation of different views of the aggregated dataflow and call graphs. Keywords—Dynamic Dataflow Graph, Function level aggregation, multi-core, program slicing that it contains all possible direct function calls, but the drawbacks are that it shows no information about the program execution and calls made through function pointers. Dynamic call graphs contain only the trace of a single execution of the software, with given input arguments. Dynamic Dataflow Graphs (DDFG) can be used to visualize the data flow of a given algorithm. DDFGs are directed acyclic graphs where nodes represents instructions and edges represents data flow. Figure 1. shows a theoretical DDFG of a four tap FIR filter calculating single output. Both ADD and MUL are arithmetic instructions, input edges are operands while output edges are results. I