--- title: TPDS 2025 Digest venue: TPDS year: 2025 date: '2025-01-01' tags: [] paper_count: 12 draft: false --- 12 papers selected. --- ### HARMONIC: Uncertainty-Aware Multi-Objective Optimization for Energy-Efficient HPC Resource Management *Kyrian Adimora, Hongyang Sun 0001* **TL;DR** — HARMONIC applies uncertainty-aware multi-objective optimization to jointly minimize energy consumption and maximize performance for HPC resource allocation. **Why notable** — It directly addresses the growing demand for energy-proportional HPC scheduling with principled probabilistic models, making it relevant to both system designers and green-computing researchers. --- ### MIST: Towards MPI Instant Startup and Termination on Tianhe HPC Systems *Yiqin Dai, Ruibo Wang, Yong Dong, Min Xie *et al.** **TL;DR** — MIST reduces MPI job startup and termination latency to near-instant on the Tianhe supercomputer by redesigning the process-management and communication-bootstrap path. **Why notable** — Startup overhead is a significant fraction of short-job turnaround time at scale; MIST's results on a top-ranked system provide a concrete reference for HPC runtime developers. --- ### Scheduling With Lightweight Predictions in Power-Constrained HPC Platforms *Danilo Carastan-Santos, Georges Da Costa, Igor Fontana De Nardin, Millian Poquet *et al.** **TL;DR** — This paper develops a scheduling framework that uses lightweight runtime predictions to respect power caps on HPC systems while minimizing job slowdown. **Why notable** — Power capping is now a first-class constraint on modern supercomputers, and this work from leading European HPC scheduling researchers offers practical, deployable algorithms. --- ### PipeMesh: Achieving Memory-Efficient Computation-Communication Overlap for Training Large Language Models *Fanxin Li, Shixiong Zhao, Yuhao Qing, Jianyu Jiang *et al.** **TL;DR** — PipeMesh overlaps pipeline-parallel computation and communication for LLM training while carefully managing memory to avoid out-of-memory failures. **Why notable** — Communication-computation overlap is one of the most impactful levers for LLM training efficiency, and PipeMesh's memory-awareness addresses the key practical constraint. --- ### EfficientMoE: Optimizing Mixture-of-Experts Model Training With Adaptive Load Balance *Yan Zeng, Chengchuang Huang, Yipeng Mei, Lifu Zhang 0004 *et al.** **TL;DR** — EfficientMoE introduces an adaptive load-balancing strategy for Mixture-of-Experts training that equalizes expert utilization and reduces communication bottlenecks. **Why notable** — MoE models are central to frontier LLM architectures, and load imbalance is their primary training inefficiency; this work provides both analysis and a practical solution. --- ### SSpMM: Efficiently Scalable SpMM Kernels Across Multiple Generations of Tensor Cores *Zeyu Xue, Mei Wen, Jianchao Yang, Minjin Tang *et al.** **TL;DR** — SSpMM delivers portable, high-performance sparse-matrix dense-matrix multiplication kernels that scale efficiently across Ampere, Hopper, and future Tensor Core generations. **Why notable** — SpMM is a bottleneck in GNN training and scientific computing; cross-generation portability without performance loss is a significant contribution for the GPU computing community. --- ### IceFrog: A Layer-Elastic Scheduling System for Deep Learning Training in GPU Clusters *Wei Gao 0064, Zhuoyuan Ouyang, Peng Sun 0006, Tianwei Zhang 0004 *et al.** **TL;DR** — IceFrog dynamically adjusts the number of pipeline stages (layers) assigned to each GPU during training to adapt to cluster heterogeneity and improve utilization. **Why notable** — Layer elasticity is a novel dimension of flexibility in distributed DNN training; IceFrog's scheduler provides measurable throughput gains in realistic heterogeneous GPU clusters. --- ### Elastic Relaxation of Concurrent Data Structures *Kåre von Geijer, Philippas Tsigas* **TL;DR** — This paper introduces a formal framework and concrete algorithms for elastic relaxation of concurrent data structures, allowing tunable trade-offs between consistency and throughput. **Why notable** — Tsigas's group advances concurrent data-structure theory with a unifying formalism that subsumes many ad-hoc relaxed designs and enables provable guarantees. --- ### Approximation Algorithms for Scheduling With/Without Deadline Constraints Where Rejection Costs are Proportional to Processing Times *Olivier Beaumont, Rémi Bouzel, Lionel Eyraud-Dubois, Esragul Korkmaz *et al.** **TL;DR** — This paper derives new approximation algorithms with tight ratios for online and offline scheduling problems where rejected jobs incur costs proportional to their processing times. **Why notable** — The theoretical results close open gaps in parallel scheduling complexity and are directly applicable to cloud and HPC batch schedulers that must handle job rejection. --- ### EdgeHydra: Fault-Tolerant Edge Data Distribution Based on Erasure Coding *Qiang He 0001, Guobiao Zhang, Jiawei Wang 0003, Ruikun Luo *et al.** **TL;DR** — EdgeHydra uses erasure coding tailored to edge-node failure patterns to provide fault-tolerant data distribution with low redundancy overhead at the network edge. **Why notable** — Fault tolerance at the edge is an increasingly critical requirement, and this system's erasure-coding approach significantly outperforms replication in storage efficiency. --- ### Two-Dimensional Balanced Partitioning and Efficient Caching for Distributed Graph Analysis *Shuai Lin, Rui Wang 0076, Yongkun Li 0001, Yinlong Xu 0001 *et al.** **TL;DR** — This paper proposes a 2D balanced graph partitioning scheme combined with a caching policy that jointly minimizes communication and replication costs in distributed graph systems. **Why notable** — Graph partitioning and caching are co-dependent problems rarely treated together; the combined optimization yields substantial performance improvements with strong theoretical backing. --- ### Towards Universal Performance Modeling for Machine Learning Training on Multi-GPU Platforms *Zhongyi Lin, Ning Sun, Pallab Bhattacharya, Xizhou Feng *et al.** **TL;DR** — This work builds a platform-agnostic performance model for distributed ML training that accurately predicts training throughput across diverse multi-GPU configurations without per-system profiling. **Why notable** — A universal modeling framework from Owens's group removes the need for expensive empirical searches when tuning distributed training configurations, benefiting the entire ML systems community.