PSSG at ACM/IEEE Supercomputing 2026

2 minute read

PSSG will have a strong presence at SC ‘26, the premier international conference for high-performance computing, networking, storage, and analysis. We are excited to be contributing in several impactful ways.

Technical Paper Presentations

Lannie Dalton Hough (PhD student) and team will be presenting their work on SpCCL: a sparsity-aware collective communication library for GPU platforms. SpCCL introduces a new bitvector-based sparse format, Pici, for fast GPU-based compression/decompression, and collective algorithms that adapt to the degree of sparsity in the data. At 99% input sparsity, SpCCL achieves up to 5.25×, 2.5×, and 2.66× speedups over NCCL for all-gather, reduce-scatter, and all-reduce, respectively. [Paper] [Talk Link]

Prof. Yafan Huang is a co-author on a paper with collaborators at the University of Florida and Argonne National Laboratory on From Error-Bounded to Fixed-Ratio: Efficient Sampling-Guided GPU Lossy Compression. This work introduces MILIO, a GPU-based lossy compression framework that combines fixed-ratio control with error-bounded compression.

Posters

Prajwal Singhania (PhD student) and Emir Gencer (PhD student) will present their poster, titled “Characterizing Scaling Performance of LLM Inference on GPU-Based Supercomputers.” Their work presents an empirical evaluation of vLLM across a range of models and workloads, comparing the scaling performance of different parallelism strategies for large-scale, multi-GPU LLM inference deployments. The poster has been selected as a candidate for the Best Research Poster Award.

Emir Gencer (PhD student) will be presenting his poster titled “Completion-Driven, Staging-Free AlltoAllv with Local-NIC Balancing”. The poster introduces an all-to-all collective communication algorithm targeted at large message sizes for MoE training. The poster demonstrates results on Frontier achieving up to 22% end-to-end training speedups. The poster has been selected as a candidate for the Best Research Poster Award.

Prajwal Singhania (PhD student) will be presenting his poster titled “Understanding and Improving Multi-Rack Symmetric-Memory AllGather in NCCL”, about modelling and benchmarking NCCL symmetric all-gather performance on NVL72 multi-rack systems. The poster also introduces improved all-gather implementations for the small-to-medium message size regimes that outperform NCCL by up to 2x.

Cunyang Wei (PhD student) will be presenting his poster titled “Plexus v2: Extreme-scale GNN Training with Communication-free Sampling and 4D Hybrid Parallelism”. The poster is about scaling mini-batch GNN training on extremely large graphs. It introduces a communication-free sampling algorithm, and combines it with 4D hybrid parallel framework. Plexus v2 scales to 2048 GPUs/GCDs on leadership-class supercomputers and achieves up to 3.5x end-to-end training speedup over state-of-the-art frameworks.

Tutorials

Daniel Nichols (former PSSG student) and Prof. Bhatele will present a tutorial titled “Rigorous Evaluation of LLM Components in HPC Research.” Attendees will learn practical techniques for evaluating LLM components in HPC research, including ensuring reproducibility, designing experiments with stochastic components, and incorporating human validation. The tutorial will also cover HPC-specific evaluation challenges, common pitfalls, and best practices for reporting results. [Session Link]

We look forward to connecting with colleagues and collaborators at SC ‘26. Stay tuned for more updates as the conference approaches!