Interactive Performance Visualization and Analysis of Execution Traces for Pattern-Based Parallel Programming

Abstract We introduce the design and implementation of a performance visualization system for high-level programming of heterogeneous parallel systems. The system consists of an execution trace extension to the algorithmic skeleton programming framework SkePU and an interactive tool that visualizes the execution trace as a graph model with a connected source code browser. The main contribution of the work lies in addressing the gap of performance analysis and debugging for programming of parallel and heterogeneous architectures at high abstraction levels, where the tool should present information at the same level of abstraction as the programming model itself. We evaluate the work by usability user surveys, observations, and interviews of novice to intermediate-level SkePU users. Additionally, we conduct performance evaluation to quantify the overhead of execution tracing of SkePU programs. Finally, we evaluate how trace lengths impact responsiveness and interactivity of the visualization tool and apply it to a larger program for training convolutional neural network models.