--- title: SoCC 2025 Digest venue: SoCC year: 2025 date: '2025-01-01' tags: [] paper_count: 13 draft: false --- 13 papers selected. --- ### From Bottleneck to Breakthrough: Optimizing Scheduling for Hyperscale Containerized Clusters *Bing Li, Yuquan Ren, Xinyi Song, Zhilei Liu *et al.** **TL;DR** — Documents production-scale scheduling improvements at a hyperscale cloud provider, demonstrating how targeted optimizations reduce scheduling tail latency and increase cluster utilization in real containerized workloads. --- ### CPU-Limits kill Performance: Time to rethink Resource Control *Chirag C. Shetty, Sarthak Chakraborty, Hubertus Franke, Larisa Shwartz *et al.** **TL;DR** — Challenges the conventional use of CPU cgroup limits in cloud environments, showing through production evidence that CFS bandwidth throttling degrades application QoS and proposing a rethink of resource control abstractions. --- ### Rethinking Tiered Memory Management in Cloud Data Centers *Tong Xing 0002, Jiaxun Yang, Javier Picorel, Antonio Barbalace* **TL;DR** — Proposes a novel tiered memory management framework for cloud data centers that improves performance by rethinking the placement and migration policies across DRAM and CXL/NVM tiers. --- ### Cost-Efficient Cloud Infrastructure with Hugepage-aware Memory Deduplication *Ruizhe Huang, Xinyu Wang 0043, Zhida An, Hanwen Lei *et al.** **TL;DR** — Deploys hugepage-aware memory deduplication in a large production cloud, achieving significant memory savings without the performance regressions that plague conventional THP-based deduplication. --- ### ALAP: Intent-Based Serverless Computing via Delayed Decision-Making *Prasoon Sinha, Kostis Kaffes, Neeraja J. Yadwadkar* **TL;DR** — Introduces an intent-based programming model for serverless that defers scheduling decisions until runtime context is available, improving resource efficiency and SLO attainment over eager placement strategies. --- ### Hydra: Virtualized Multi-Language Runtime for High-Density Serverless Platforms *Serhii Ivanenko, Vasyl Lanko, Rudi Horn, Vojin Jovanovic *et al.** **TL;DR** — Presents a virtualized runtime that multiplexes multiple language environments within a single sandbox, enabling higher function density and faster cold starts on serverless platforms. --- ### Serverless Elasticsearch: the Architecture Transformation from Stateful to Stateless *Iraklis Psaroudakis, Pooya Salehi, Jason Bryan, Francisco Fernández Castaño *et al.** **TL;DR** — Describes Elastic's production migration of Elasticsearch to a serverless, stateless architecture, sharing engineering lessons on decoupling compute from state at cloud scale. --- ### DFUSE: Strongly Consistent Write-Back Kernel Caching for Distributed Userspace File Systems *Haoyu Li, Jingkai Fu, Qing Li 0002, Windsor Hsu *et al.** **TL;DR** — Closes a long-standing gap in FUSE-based distributed file systems by enabling strongly consistent write-back caching in the kernel, significantly improving throughput without sacrificing correctness. --- ### Accelerating Distributed Filesystem Metadata Service via Decoupling Directory Semantics from Metadata Indexing *Wenhao Lv, Hao Guo, Qing Wang 0031, Youyou Lu *et al.** **TL;DR** — Achieves scalable distributed filesystem metadata by separating directory namespace semantics from the underlying index structure, reducing contention and improving throughput for large-scale cloud storage. --- ### Valet: Efficient Data Placement on Modern SSDs *Devashish R. Purandare, Peter Alvaro, Avani Wildani, Darrell D. E. Long *et al.** **TL;DR** — Exploits fine-grained internal SSD geometry to make smarter data placement decisions, yielding measurable I/O performance gains without changes to the host storage stack. --- ### Understanding Diffusion Model Serving in Production: A Top-Down Analysis of Workload, Scheduling, and Resource Efficiency *Yanying Lin, Shuaipeng Wu, Shutian Luo, Hong Xu 0001 *et al.** **TL;DR** — Provides the first comprehensive production characterization of diffusion model inference workloads, revealing unique scheduling and resource efficiency challenges distinct from LLM serving. --- ### ModServe: Modality- and Stage-Aware Resource Disaggregation for Scalable Multimodal Model Serving *Haoran Qiu, Anish Biswas, Zihan Zhao, Jayashree Mohan *et al.** **TL;DR** — Disaggregates compute resources per modality and pipeline stage for multimodal inference, with a Microsoft production deployment showing improved GPU utilization and latency over monolithic serving. --- ### THORN-ML: Transparent Hardware Offloaded Resilient Networks for RDMA based Distributed ML Workloads *Maziyar Nazari, Daniel Noland, Giulio Sidoretti, Erika Hunhoff *et al.** **TL;DR** — Offloads RDMA fault detection and recovery to programmable network hardware, making distributed ML training resilient to network failures without modifying the training framework or incurring software overhead.