Hans Meuer Award Session

This session presents the Hans Meuer Award for the best research paper at ISC 2026, chaired by Hatem Ltaief of KAUST, who reports 125 full submissions (a 30 percent increase), 132 program committee members, and a 27 percent acceptance rate. The award goes to PICO: Performance Insight for Collective Operations, presented by Saverio Pasqualoni with collaborators from Sapienza University of Rome, KAUST, and ETH Zurich, including Marco Canini and Torsten Hoefler. The talk motivates why collective communication tuning is hard: collectives can dominate MPI runtime on production supercomputers and consume up to a third of mixture-of-experts training time, while heterogeneous networks, dragonfly topologies, node allocation effects, and layered software stacks (NCCL, MPI, UCC/UCX, libfabrics) expose dozens to hundreds of tuning parameters. A binomial broadcast example shows how distance-halving versus distance-doubling variants with identical cost models can differ 2.5x in practice due to global-link traffic. PICO moves from end-to-end benchmarking to fine-grained diagnosis by tagging semantically meaningful algorithm phases, decoupling test descriptors from platform descriptors for portability, and capturing extensive metadata for reproducibility, with post-processing tools including a network congestion tracer and tuning file generator. Results on LUMI show default all-reduce algorithm choices leaving substantial performance unrealized, and trace-driven simulation of Llama and Mixtral training shows up to 44 percent faster iteration time from changing collective algorithms alone. Questions address non-blocking collectives and reproducibility under nondeterminism.