Vincent Lannurien
d822cdaa6a
6.8 KiB
title, venue, year, date, tags, paper_count, draft
| title | venue | year | date | tags | paper_count | draft |
|---|---|---|---|---|---|---|
| FGCS 2025 Digest | FGCS | 2025 | 2025-01-01 | 12 | false |
12 papers selected.
Multifacets of lossy compression for scientific data in the Joint-Laboratory of Extreme Scale Computing
Franck Cappello, Mario C. Acosta, Emmanuel Agullo, Hartwig Anzt et al.
TL;DR — A joint JLESC survey covering error-bounded lossy compressors (SZ, ZFP, MGARD) across simulation, AI, and in-situ analytics use cases, with benchmarks on real scientific datasets at extreme scale.
Why notable — The most comprehensive cross-site evaluation of scientific data compression to date, providing actionable guidance on compressor selection for different numerical kernels and accuracy requirements.
Efficient distributed continual learning for steering experiments in real-time
Thomas Bouvier, Bogdan Nicolae, Alexandru Costan, Tekin Bicer et al.
TL;DR — Proposes a distributed continual learning architecture that keeps deep learning models synchronized with a running scientific experiment by streaming lightweight updates across edge detectors and HPC backends.
Why notable — One of the first systems to close the loop between experimental data streams and model adaptation in real-time without full retraining, validated on synchrotron detector workloads.
SmartKV: A cost-effective and low-latency geo-distributed key-value store for the computing continuum
Juan Aznar-Poveda, Maximilian Franz Ebner, Thomas Fahringer, Zahra Najafabadi Samani et al.
TL;DR — Introduces SmartKV, a geo-distributed key-value store that uses latency-aware replication policies to deliver consistent low-latency reads across edge, fog, and cloud tiers of the computing continuum.
Why notable — Provides a concrete, benchmarked storage primitive for the computing continuum that fills the gap between single-datacenter stores and high-latency cloud object storage.
Scalable compute continuum
Valeria Cardellini, Patrizio Dazzi, Gabriele Mencagli, Matteo Nardelli 0001 et al.
TL;DR — Defines a programming and deployment model for the compute continuum that abstracts resource heterogeneity from edge to cloud, enabling applications to scale dynamically across tiers.
Why notable — Provides a principled architectural reference for the continuum that can guide system designers building next-generation distributed runtime environments.
A comparative study of ad-hoc file systems for extreme scale computing
Njoud O. Almaaitah, Francisco Javier García Blas, Genaro Sanchez-Gallegos, Jesús Carretero 0001 et al.
TL;DR — Benchmarks GekkoFS, BeeGFS, and similar ad-hoc file systems under diverse HPC I/O patterns, characterizing their throughput, metadata performance, and suitability for burst-buffer scenarios.
Why notable — The most systematic evaluation of ad-hoc parallel file systems available, giving HPC centers clear data to choose or configure temporary storage for large scientific workflows.
Advancing anomaly detection in computational workflows with active learning
Krishnan Raghavan, George Papadimitriou 0002, Hongwei Jin, Anirban Mandal et al.
TL;DR — Applies active learning to reduce the labeling burden for workflow anomaly detection, selectively querying an oracle for the most informative execution traces within a Pegasus workflow framework.
Why notable — Demonstrates that active learning can make anomaly detection practical in real scientific workflows where labeled failure data is scarce, with experiments on production workloads.
MITgcm-AD v2: Open source tangent linear and adjoint modeling framework for the oceans and atmosphere enabled by the Automatic Differentiation tool Tapenade
Shreyas Sunil Gaikwad, Sri Hari Krishna Narayanan, Laurent Hascoët, Jean-Michel Campin et al.
TL;DR — Describes MITgcm-AD v2, a production-quality adjoint of the MITgcm ocean-atmosphere model generated with Tapenade, enabling global sensitivity analyses and data assimilation at scale.
Why notable — A landmark in scientific computing software sustainability: a fully open, differentiable climate model that enables gradient-based inversion for ocean state estimation.
zCeph: Design and implementation of a ZNS-friendly distributed file system
Jinyong Ha 0001, Yongseok Son
TL;DR — Redesigns the Ceph distributed file system to exploit Zoned Namespace SSDs natively, eliminating write amplification and improving throughput by aligning file system semantics with ZNS zone constraints.
Why notable — Demonstrates how next-generation storage hardware (ZNS SSDs) demands rethinking distributed storage stack designs, with significant performance gains on real hardware.
RADiCe: A Risk Analysis Framework for Data Centers
Fabian Mastenbroek, Tiziano De Matteis, Vincent van Beek, Alexandru Iosup
TL;DR — Provides a quantitative risk analysis framework for data centers that models cascading failures across power, cooling, and compute subsystems using simulation to estimate availability and cost trade-offs.
Why notable — Fills a practical gap for data center operators who need principled tools to evaluate infrastructure resilience beyond simple redundancy rules.
Deadline-constrained security-aware workflow scheduling in hybrid cloud architecture
Somayeh Abdi, Mohammad Ashjaei, Saad Mubeen
TL;DR — Formulates workflow scheduling in hybrid clouds as a multi-objective problem that jointly minimizes cost and execution time while meeting both deadline and data-security placement constraints.
Why notable — One of the few scheduling works that treats security classification of tasks as a first-class constraint alongside performance, with practical validation on scientific workflow benchmarks.
Regen: An object layout regenerator on large-scale production HPC systems
Dong Kyu Sung, Sunggon Kim, Sangjin Lee 0003, Houjun Tang et al.
TL;DR — Regen transparently reorganizes the on-disk layout of HDF5 and NetCDF objects in parallel file systems to match actual access patterns, improving I/O performance without application changes.
Why notable — Deployed and validated on a production HPC system, showing significant I/O speedups for real scientific datasets, making it immediately relevant to storage administrators.
Formal definition and implementation of reproducibility tenets for computational workflows
Nicholas J. Pritchard, Andreas Wicenec
TL;DR — Formalizes a set of reproducibility requirements for scientific workflows and implements a verification layer within the DALIUGE workflow engine that checks compliance at design and execution time.
Why notable — Provides the community with a concrete, tool-supported definition of workflow reproducibility, moving beyond aspirational guidelines to enforceable runtime checks.