Vincent Lannurien
d822cdaa6a
6.5 KiB
title, venue, year, date, tags, paper_count, draft
| title | venue | year | date | tags | paper_count | draft |
|---|---|---|---|---|---|---|
| JPDC 2025 Digest | JPDC | 2025 | 2025-01-01 | 12 | false |
12 papers selected.
Throughput of Byzantine Broadcast
Ruomu Hou, Haifeng Yu, Prateek Saxena
TL;DR — Establishes tight throughput bounds for Byzantine broadcast protocols and constructs algorithms that saturate those bounds, separating throughput from latency in the fault-tolerant broadcast landscape.
Why notable — Provides the first rigorous throughput characterization of Byzantine broadcast, a fundamental primitive whose capacity limits were previously unquantified.
How to reduce the number of steps for (multi-valued validated) Byzantine agreement?
Baohan Huang, Haibin Zhang, Chao Liu 0039, Shengli Liu 0001 et al.
TL;DR — Presents new Byzantine agreement protocols that lower the step complexity for multi-valued and validated variants, breaking barriers that have stood since the classical results.
Why notable — Step complexity is a fundamental metric for distributed agreement; reducing it has direct implications for consensus latency in blockchains and replicated systems.
Locating a black hole in a dynamic ring
Giuseppe Antonio Di Luna, Paola Flocchini, Giuseppe Prencipe, Nicola Santoro
TL;DR — Solves the black-hole search problem on rings whose topology changes over time, establishing the agent and time complexity of locating a fatal node in a dynamic distributed environment.
Why notable — Extends a classic distributed exploration problem to dynamic graphs, requiring new algorithmic techniques that are broadly applicable to fault detection in evolving networks.
Dispersion of mobile robots on directed anonymous graphs
Giuseppe F. Italiano, Debasish Pattanayak, Gokarna Sharma
TL;DR — Characterizes the necessary and sufficient conditions for a group of mobile robots to disperse to distinct nodes of a directed anonymous graph, and provides optimal algorithms.
Why notable — Directed anonymous graphs model asymmetric communication networks; the dispersion problem's resolution here advances the theory of autonomous distributed agents.
QPOPSS: Query and Parallelism Optimized Space-Saving for finding frequent stream elements
Victor Jarlow, Charalampos Stylianopoulos, Marina Papatriantafilou
TL;DR — Redesigns the Space-Saving frequent-elements sketch for concurrent shared-memory execution, achieving high query throughput alongside update throughput without sacrificing approximation accuracy.
Why notable — Bridges the gap between approximate streaming data structures and parallel execution, demonstrating that heavy-hitter summaries can scale on multicore without significant accuracy loss.
A parallel algorithm for minimum weight set cover with small neighborhood property
Yingli Ran, Yaoyao Zhang, Zhao Zhang 0002
TL;DR — Gives a parallel approximation algorithm for minimum weight set cover instances where sets have bounded neighborhood size, achieving near-optimal approximation ratio in poly-logarithmic rounds.
Why notable — Expands the frontier of problems admitting efficient parallel approximation, with implications for distributed network optimization where local structure can be exploited.
Optimizing parallel heterogeneous system efficiency: Dynamic task graph adaptation with recursive tasks
Nathalie Furmento, Abdou Guermouche, Gwenolé Lucas, Thomas Morin et al.
TL;DR — Extends task-graph runtime systems to support recursive task generation, enabling dynamic adaptation of the task graph structure to improve load balance on heterogeneous CPU-GPU platforms.
Why notable — Recursive task parallelism is essential for divide-and-conquer workloads; integrating it into heterogeneous runtimes closes a major gap in practical parallel programming models.
A scheduler to foster data locality for GPU and out-of-core task-based linear algebra applications
Maxime Gonthier, Loris Marchal, Samuel Thibault
TL;DR — Proposes a data-locality-aware scheduler for task-based dense linear algebra that simultaneously manages GPU memory and out-of-core data transfers to minimize data movement.
Why notable — Data movement dominates cost in large linear algebra computations; the scheduler's dual handling of GPU memory and disk I/O makes it practically relevant for exascale workloads.
Leveraging Multi-Instance GPUs through moldable task scheduling
Jorge Villarrubia, Luis Costero, Francisco D. Igual, Katzalin Olcoz
TL;DR — Develops a moldable task scheduling framework that dynamically partitions GPU compute across concurrent tasks using NVIDIA's Multi-Instance GPU feature to improve overall throughput.
Why notable — MIG is a critical hardware feature for multi-tenant GPU clusters; this work provides the first scheduling framework that exploits it through principled moldable-task theory.
Integration framework for online thread throttling with thread and page mapping on NUMA systems
Janaina Schwarzrock, Hiago Mayk G. de A. Rocha, Arthur Francisco Lorenzon, Samuel Xavier de Souza et al.
TL;DR — Combines online thread-count throttling with NUMA-aware thread and page placement in a unified runtime framework, adaptively co-optimizing both dimensions to maximize performance.
Why notable — Thread throttling and NUMA placement are typically managed independently; their joint online optimization yields measurable gains that neither technique alone achieves.
To repair or not to repair: Assessing fault resilience in MPI stencil applications
Roberto Rocco, Elisabetta Boella, Daniele Gregori, Gianluca Palermo
TL;DR — Systematically evaluates the cost-benefit trade-off between full fault recovery and partial resilience strategies for MPI stencil computations under process failures.
Why notable — Provides practitioners with a principled decision framework for resilience in HPC applications, showing when expensive full recovery is justified versus cheaper degraded-mode execution.
A lightweight RDMA connection protocol based on post-hoc confirmation
Ke Wu 0003, Dezun Dong, Weixia Xu 0001
TL;DR — Designs an RDMA connection protocol that defers acknowledgment to post-operation confirmation, drastically reducing connection setup overhead for short-lived high-frequency transfers.
Why notable — RDMA setup latency is a critical bottleneck in disaggregated memory and distributed storage systems; this protocol's approach generalizes to any latency-sensitive fabric.