Birds of a Feather Archive

HPC Graph Toolkits and the GraphBLAS Forum


Authors: Antonino Tumeo (Pacific Northwest National Laboratory (PNNL)), Timothy Mattson (Merly.ai), Jose Moreira (IBM Thomas J. Watson Research Center)

Abstract: Graph analytics is foundational to scientific discovery, national security, artificial intelligence (AI), and industrial innovation, yet sparse, irregular, and evolving graph workloads continue to challenge HPC systems. This BoF convenes developers and users of high-performance graph toolkits to exchange updates, compare combinatorial and algebraic approaches, and define priorities for converged workflows spanning simulation, data analytics, AI, instruments, edge platforms, clouds, and supercomputers. Discussion will address GraphBLAS, interoperability, dynamic graphs, heterogeneous execution, provenance, trustworthy reasoning, and sustainable open infrastructure, producing concrete community directions for research, software, standards, benchmarks, infrastructure investment, and cross-sector collaboration. The outcome will guide collaborative work beyond SC26.

Long Description: Graphs are a fundamental abstraction for representing relationships, dependencies, and evolving structure in scientific, national-security, and industrial data. They underpin molecular and materials modeling, power and transportation networks, cybersecurity, intelligence analysis, logistics, digital twins, knowledge systems, and AI. Their sparse, irregular, attributed, and often dynamic behavior, however, continues to expose limitations in HPC algorithms, architectures, runtimes, programming models, and data-management systems. Held at SC since 2017, the HPC Graph Toolkits and GraphBLAS Forum has become a sustained, vendor-neutral community meeting, regularly attracting more than 100 participants from academia, national laboratories, government, and industry. It brings together developers and users of algebraic, combinatorial, task-based, and asynchronous graph systems to exchange progress, compare approaches, and identify challenges that no single toolkit or project can address. The discussion spans GraphBLAS and sparse linear algebra, graph algorithms and databases, distributed runtimes, accelerator support, and mature and emerging ecosystems such as SuiteSparse, LAGraph, CombBLAS, GraphIt, Galois, Gunrock, cuGraph, and SHAD. A central focus for 2026 is the role of graphs in converged workflows combining scientific simulation, experimental observations, data analytics, and AI. These workflows increasingly form closed loops: instruments and simulations generate data; graph analytics reveal relationships, anomalies, and evolving structures; AI models infer or rank candidate outcomes; and results steer subsequent simulations, experiments, or data acquisition. They may span sensors and instruments, edge and laboratory systems, cloud services, data facilities, and leadership-class supercomputers. Graphs can serve not only as scientific data structures, but also as representations of workflow dependencies, provenance, knowledge, resources, and relationships among models and datasets. The DOE Genesis Mission’s American Science and Security Platform provides a timely example of the need to integrate advanced computing, AI, scientific facilities, and distributed data resources. Related needs arise across DOE and NNSA; defense and intelligence organizations such as DARPA and IARPA; civil agencies including NSF, NASA, and DHS; and industry. The BoF will examine how high-performance graph infrastructure can support these environments without limiting the agenda to any single mission, agency, or application domain. Open research questions include incremental and streaming analysis of evolving, temporal, and heterogeneous graphs; distributed and asynchronous GraphBLAS; interoperability among graph toolkits, databases, AI frameworks, workflow systems, and simulation codes; graph-aware placement of data and computation across edge, cloud, and HPC resources; secure cross-facility execution; communication avoidance, resilience, and performance portability; provenance, uncertainty, explainability, and trustworthy reasoning; and end-to-end benchmarks measuring time to scientific or mission insight rather than isolated kernel performance. Particular attention will be given to whether common representations and composable interfaces can preserve graph structure and semantics as data moves among databases, analytics, graph neural networks (GNNs), knowledge graphs, simulations, and workflow services, avoiding costly conversions. The session will also consider standards, sustainable open-source infrastructure, accelerator-aware algorithms, and capabilities requiring long-term public-private coordination beyond isolated commercial offerings. The expected outcome is a community-driven research and development agenda identifying shared abstractions, interfaces, benchmarks, infrastructure needs, and concrete opportunities for collaboration across algorithms, systems, architectures, applications, agencies, and industry.

Website: https://hpc.pnl.gov/BOF/



Back to Birds of a Feather Archive Listing