Work Visualizer: A User Support Tool to Facilitate General Computing Using Heterogeneous Architectures at Scale

One of the main obstacles to the propagation of ASCR-funded high-performance computing (HPC) libraries (e.g, MPICH, OpenMPI, Kokkos, Legion, RAJA) is the challenge posed by performance optimization of application codes that run on heterogeneous architectures at scale. A visual, interactive and intuitive tool to guide users through the execution of the application and highlight areas of interest, thereby helping to save both developer and compute time, as well as electricity, is thus of utmost importance not only to DOE but also to the commercial “tech” sector at-large. We proposed to develop the “WorkVisualizer,” a browser-based application integrating informative visualization to coarsen up a level of abstraction over the fine granularity of trace-based profiling (e.g. like the Projections tool) and support users who run at large scale. In Phase I, we concretized the design and built the WorkVisualizer on three pillars: (a) Collection: leverages Caliper (Lawrence Livermore National Laboratory) via a custom configuration to collect program data without requiring changes to source code for MPI+Kokkos apps. (b) Analysis: calculates high-level summary information and cross-rank time segmented information to determine how the application can be optimized. (c) Visualization: provides a variety of interactive visualizations that offer actionable insights into the structure and execution patterns of the application. In particular, we have developed and made public an open-source proof-of-concept of the proposed functionality tested on small-scale, CPU-only data up to a few hundred ranks. In Phase II, we will focus on improving the scalability of the WorkVisualizer while generalizing the three pillars to apply to a broader range of scientific programs, by: (1) accelerating the time to results, (2) adding support for GPU devices, (3) improving upon the existing analytical metrics and algorithms, and (4) increasing the interactivity and scope of the visualizations. These efforts will culminate in an improved version of the WorkVisualizer that can quickly analyze and visualize HPC applications that run at large scale on heterogeneous architectures, all while maintaining the ease-of-use and lightweight functionality that we established in Phase I. By facilitating the performance optimization of high-performance libraries on heterogeneous computing platform, the WorkVisualizer could lead to faster advancements in key scientific and engineering domains. With a commitment to open-source scientific software, our commercialization strategy includes (1) providing a tiered licensing offering with the WorkVisualizer as a Software-as-a-Service (SaaS) product, from the individual developer to the high-performance computing center administrator, and (2) support and customization services.