--- title: EuroSys 2025 Paper Summaries and Analysis | daily.dev description: A comprehensive analysis of over 40 papers from EuroSys 2025, covering AI/ML systems, cloud computing, networking, storage, and security. Key themes include:... canonical: https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: EuroSys 2025 Paper Summaries and Analysis | daily.dev og:description: A comprehensive analysis of over 40 papers from EuroSys 2025, covering AI/ML systems, cloud computing, networking, storage, and security. Key themes include:... og:url: https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o og:image: https://api.daily.dev/og/posts/pg5AxGo2O.png og:image:alt: EuroSys 2025 Paper Summaries and Analysis og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # EuroSys 2025 Paper Summaries and Analysis **[eunomia eBPF Blog](https://daily.dev/sources/eunomia)** · 58 min read · 0 upvotes · 0 comments ## Summary A comprehensive analysis of over 40 papers from EuroSys 2025, covering AI/ML systems, cloud computing, networking, storage, and security. Key themes include: bespoke systems for LLM training and inference (Mist, MEPipe, HybridFlow, HCache, CacheBlend, T-MAC, DeltaZip), resource disaggregation and CXL-attached memory, congestion control innovations for datacenters, GPU sharing and scheduling, serverless cold starts, BFT consensus improvements using TEEs, and network verification at scale. The analysis identifies cross-stack hardware-software co-design, data-driven optimization replacing static heuristics, and proactive rather than reactive system design as dominant methodological trends. Open challenges include managing system complexity, comprehensive performance isolation across all resource types, and verification tools for ML and disaggregated systems. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What speedup does Mist achieve over Megatron-LM for LLM distributed training? Mist achieves up to 1.73× faster training than Megatron-LM used as a manual baseline, and up to 2.04× speedup over the prior automation tool Aceso. It co-optimizes memory reduction techniques and parallelism strategies through overlap-centric scheduling and symbolic performance modeling, automatically finding efficient configurations for data, tensor, and pipeline parallelism. _Teams benchmarking LLM training frameworks track results like these on daily.dev before committing to an infrastructure choice._ ### How does T-MAC speed up LLM inference on CPUs compared to llama.cpp? T-MAC achieves up to 4× higher throughput and roughly 70% less energy use than llama.cpp on CPUs. It replaces matrix multiplications on quantized models with bit-wise table lookups, precomputing partial results for low-bit weight and activation combinations. On an Apple M2 Ultra it generates 71 tokens/sec (8 cores) for a 3B model; a Raspberry Pi 5 reaches 11 tokens/sec. _Engineers deploying LLMs on edge or CPU-only hardware follow developments like T-MAC on daily.dev._ ### What throughput improvement does DeltaZip provide for serving multiple fine-tuned LLMs? DeltaZip improves throughput by 2× to 12× compared to naive per-model deployment. It compresses only the weight deltas between fine-tuned variants and a shared base model, shrinking fine-tune updates by up to 10× while preserving accuracy, then applies compressed deltas on the fly at query time in a multi-tenant serving framework. _Engineers running multi-tenant LLM serving infrastructure keep up with efficiency techniques like DeltaZip on daily.dev._ ## Similar posts on daily.dev - [Best of metadata in 2025](https://daily.dev/posts/best-of-metadata-in-2025-gs9xv613f) · Metadata · 1 upvotes · 0 comments - [Data in 2026: Interchangeable Models, Clouds, and Specialization](https://daily.dev/posts/data-in-2026-interchangeable-models-clouds-and-specialization-wdn2eeq5x) · The New Stack · 2 upvotes · 0 comments - [2025 Key Trends: AI Workflows, Architectural Complexity, Sociotechnical Systems & Platform Products](https://daily.dev/posts/2025-key-trends-ai-workflows-architectural-complexity-sociotechnical-systems-platform-products-hylunyae4) · InfoQ · 0 upvotes · 0 comments - [2025 Recap: The Year Software Development Changed Shape](https://daily.dev/posts/2025-recap-the-year-software-development-changed-shape-7p7qgtf7j) · Docker · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#cloud](https://daily.dev/tags/cloud), [#llm](https://daily.dev/tags/llm), [#distributed-systems](https://daily.dev/tags/distributed-systems) [View this post on daily.dev](https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"EuroSys 2025 Paper Summaries and Analysis","url":"https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o"},"datePublished":"2026-08-11T08:54:06.165Z","dateModified":"2026-08-11T09:09:31.030Z","description":"A comprehensive analysis of over 40 papers from EuroSys 2025, covering AI/ML systems, cloud computing, networking, storage, and security. Key themes include:...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b59e04eefe0ea85df0b7e50151b0b12e?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/b59e04eefe0ea85df0b7e50151b0b12e?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"eunomia eBPF Blog","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"eunomia eBPF Blog","logo":"https://media.daily.dev/image/upload/s--z-nTl7TK--/f_auto,q_auto/v1786438304/logos/eunomia?_a=BAMAMicg0","url":"https://daily.dev/sources/eunomia"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,cloud,llm,distributed-systems","timeRequired":"PT58M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"eunomia eBPF Blog","item":"https://daily.dev/sources/eunomia"},{"@type":"ListItem","position":3,"name":"EuroSys 2025 Paper Summaries and Analysis"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/eurosys-2025-paper-summaries-and-analysis-pg5axgo2o#faq","mainEntity":[{"@type":"Question","name":"What speedup does Mist achieve over Megatron-LM for LLM distributed training?","acceptedAnswer":{"@type":"Answer","text":"Mist achieves up to 1.73× faster training than Megatron-LM used as a manual baseline, and up to 2.04× speedup over the prior automation tool Aceso. It co-optimizes memory reduction techniques and parallelism strategies through overlap-centric scheduling and symbolic performance modeling, automatically finding efficient configurations for data, tensor, and pipeline parallelism. Teams benchmarking LLM training frameworks track results like these on daily.dev before committing to an infrastructure choice."}},{"@type":"Question","name":"How does T-MAC speed up LLM inference on CPUs compared to llama.cpp?","acceptedAnswer":{"@type":"Answer","text":"T-MAC achieves up to 4× higher throughput and roughly 70% less energy use than llama.cpp on CPUs. It replaces matrix multiplications on quantized models with bit-wise table lookups, precomputing partial results for low-bit weight and activation combinations. On an Apple M2 Ultra it generates 71 tokens/sec (8 cores) for a 3B model; a Raspberry Pi 5 reaches 11 tokens/sec. Engineers deploying LLMs on edge or CPU-only hardware follow developments like T-MAC on daily.dev."}},{"@type":"Question","name":"What throughput improvement does DeltaZip provide for serving multiple fine-tuned LLMs?","acceptedAnswer":{"@type":"Answer","text":"DeltaZip improves throughput by 2× to 12× compared to naive per-model deployment. It compresses only the weight deltas between fine-tuned variants and a shared base model, shrinking fine-tune updates by up to 10× while preserving accuracy, then applies compressed deltas on the fly at query time in a multi-tenant serving framework. Engineers running multi-tenant LLM serving infrastructure keep up with efficiency techniques like DeltaZip on daily.dev."}}]} ```