Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms
Authors: Ahmedur Rahman Shovon and Romain Pereira (Argonne National Laboratory), Yihao Sun (Utah State University), Farhan Saif and Sidharth Kumar (University of Illinois Chicago), and Swann Perarnau (Argonne National Laboratory)
Abstract: Recursive Datalog queries such as transitive closure, connected components, and same generation iterate a rule to a fixpoint and express many graph analytics. GPU Datalog engines run each round as a chain of separate relational operators driven by a host loop, paying for repeated sorting, per-round allocation, and a CPU-GPU round-trip. Instead, we run each round as single CUDA kernel over GPU-resident data structure, so a single atomic operation per derived fact deduplicates, tests novelty, and inserts it together. Because the state stays on the device, the whole loop runs on the GPU via a replayed CUDA graph and a conditional graph node, and results are updated in place as edges arrive. On an NVIDIA A100, this cuts end-to-end time by up to 24x for transitive closure and 136x for connected components over MNMGDatalog, beats GPUlog by up to 29x in compute, and updates connected components up to 40x faster.
Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing
Back to Full Workshop Archive Listing