Vincent Lannurien
d822cdaa6a
6.6 KiB
title, venue, year, date, tags, paper_count, draft
| title | venue | year | date | tags | paper_count | draft | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HPDC 2025 Digest | HPDC | 2025 | 2025-07-20 |
|
10 | false |
10 papers selected.
Parameterized Algorithms for Non-uniform All-to-all
Ke Fan, Jens Domke, Seydou Ba, Sidharth Kumar
TL;DR — Introduces parameterized algorithms that adapt non-uniform all-to-all collective communication to heterogeneous network topologies, reducing message contention and improving throughput.
Why notable — Non-uniform all-to-all is a performance bottleneck in many HPC applications; topology-aware parameterization directly benefits MPI implementations on dragonfly and fat-tree networks at scale.
DPU-KV: On the Benefits of DPU Offloading for In-Memory Key-Value Stores at the Edge
Arjun Kashyap, Yuke Li 0003, Xiaoyi Lu 0001
TL;DR — Offloads key-value store operations to Data Processing Units (DPUs) over RDMA to reduce CPU overhead and tail latency in edge deployments.
Why notable — DPU offloading is an emerging paradigm for network-attached smart NICs in HPC clusters; this work provides concrete performance analysis showing when and how much offloading helps, informing future RDMA-based storage designs.
TSUE: A Two-Stage Data Update Method for an Erasure Coded Cluster File System
Zheng Wei, Jing Xing, Yida Gu, Wenjing Huang 0002 et al.
TL;DR — Proposes a two-stage update scheme for erasure-coded parallel file systems that decouples the logging and parity-update phases to cut write amplification and I/O latency.
Why notable — Erasure coding is increasingly used in large-scale HPC storage to replace replication, but update overhead remains a bottleneck; TSUE addresses a core pain point for Lustre- and GPFS-class parallel file systems.
LegoIndex: A Scalable and Modular Indexing Framework for Efficient Analysis of Extreme-Scale Particle Data
Chang Guo, Ning Yan 0002, Lipeng Wan 0001, Zhichao Cao 0002
TL;DR — Presents a composable, multi-level indexing framework for particle simulation datasets that enables efficient query processing at extreme scale without requiring full dataset scans.
Why notable — Scientific particle simulations at exascale generate data volumes that overwhelm traditional post-processing pipelines; LegoIndex's modular design allows it to be adapted across different storage backends and query patterns commonly seen in DOE workloads.
IPComp: Interpolation Based Progressive Lossy Compression for Scientific Applications
Zhuoxun Yang, Sheng Di, Longtao Zhang, Ruoyu Li et al.
TL;DR — Introduces interpolation-driven progressive lossy compression that lets users trade accuracy for compression ratio at query time rather than at write time, without re-compressing stored data.
Why notable — Progressive reconstruction is a long-sought capability for HPC I/O; IPComp achieves it with competitive compression ratios and builds on the widely used SZ/ZFP lineage, making adoption in existing scientific workflows straightforward.
Advancing Scientific Data Compression via Cross-Field Prediction
Youyuan Liu, Wenqi Jia 0003, Taolue Yang, Bo Jiang et al.
TL;DR — Exploits correlations between different physical fields in multi-field scientific datasets to improve lossy compression ratios beyond what single-field methods can achieve.
Why notable — Multi-field simulations (climate, combustion, fusion) dominate HPC storage consumption; cross-field prediction represents a principled, generally applicable step change in compression efficiency for these workloads.
Flux Emulator: First Insights into Optimizing Scheduling for Exascale HPC
W. Jay Ashworth, Ian Lumsden, Jim Garlick, Mark Grondona et al.
TL;DR — Presents an emulation infrastructure for the Flux workload manager that enables scheduling algorithm evaluation at exascale node counts without requiring access to a full exascale machine.
Why notable — Validating schedulers at exascale is otherwise infeasible before systems exist; Flux Emulator directly supports the scheduling research needed to maximize utilization of Frontier- and Aurora-class systems.
HYPERF: End-to-End Autotuning Framework for High-Performance Computing
Juseong Park, Yongwon Shin, Junghyun Lee, Junseo Lee et al.
TL;DR — Delivers an end-to-end autotuning framework that jointly optimizes compiler flags, runtime parameters, and problem-specific configurations for HPC applications through structured search.
Why notable — Manual tuning of HPC codes for new architectures is expensive and error-prone; HYPERF's end-to-end scope distinguishes it from prior tools that target only one layer of the software stack, offering broader applicability across the HPC software ecosystem.
Efficient and Cost-Effective HPC on the Cloud
Aditya Bhosale, Laxmikant V. Kalé, Sara Kokkila Schumacher
TL;DR — Demonstrates how Charm++-based adaptive runtime techniques—load balancing, dynamic over-decomposition, and message-driven execution—can recover near-on-premises HPC performance on cloud instances despite higher network variability.
Why notable — Cloud-HPC convergence is a major community priority as on-premises clusters face procurement delays; this paper provides a practitioner-oriented analysis of which runtime adaptations deliver the best performance-per-dollar on AWS and Azure.
Bringing Differential Privacy to HPC: Privacy-Preserving Transformations of HPC Traces
Ana Luisa Veroneze Solórzano, Rohan Basu Roy, Benjamin Schwaller, Sara Petra Walton et al.
TL;DR — Applies differential privacy mechanisms to HPC job and performance traces, enabling centers to share workload data for research without exposing sensitive user or application information.
Why notable — Sharing HPC traces is critical for reproducible scheduling and performance research but is often blocked by privacy concerns; this work provides a rigorous, deployable solution that could unlock a significant new supply of public HPC datasets.