1. KV Cache管理架构演进：从连续分配到统一混合内存架构 - deephub https://zhuanlan.zhihu.com/p/2012270020915389042 2. 统一的AI媒体生产线 - wangleineo的 https://zhuanlan.zhihu.com/p/1948887920762157033 3. Deepseek-v4 引入 MHC 架构实现性能稳健，如何理解其数学原理与大模型发展影响？ - PP19 https://www.zhihu.com/question/1991253856454218680/answer/1992715126378349349 4. NVIDIA Fabric Manager 架构设计、工程实现与处理流程全解析 - 码夫祥子…
1. PoWER Never Corrupts: Tool-Agnostic Verification of Crash Consistency and Corruption Detection [OSDI 2025 Best paper, 6 refs, Azure Storage, Hayley LeBlanc] https://www.microsoft.com/en-us/research/publication/power-never-corrupts-tool-agnostic-verification-of-crash-consistency-and-corruption-detection/ https://people.csail.mit.edu/nickolai/papers/leblanc-power.pdf…
1. DeepTech深科技: 谷歌AI十年棋局：从猫识别到太空数据中心，Gemini如何统一生态 https://zhuanlan.zhihu.com/p/1977401514071959136 1. Useful reference as how Google organized its AI strategy company-wide that eventually led to success 2. 2025 AI 记忆系统大横评：从插件到操作系统，谁在定义下一代Agent Infra？ - 机器之心 https://zhuanlan.zhihu.com/p/1978869876396413893 1. Good. 2. Highlights 1. 我们梳理了2024–2025年全球开源社区出现的十几种完全不同的 Memory 技术路线和代表项目： 自动抽取记忆（Mem0） 企业级…
This is a machine translation of LLM Fundamentals - Visualizing Transformer Internals CN The iteration of large models (Large Language Model, LLM) technology has been like a technological explosion; techniques that were cutting-edge six months ago have now become the basis of the basics. Polo Club created an animated explainer of the Transformer model, which is very helpful for visualizing and…
大模型（Large Language Model ，LLM）技术的迭代犹如技术爆炸，半年前的先进技术如今已成基础的基础。Polo Club 做了 Transformer 模型的动态图解，对可视化地理解 Transformer 的工作原理非常友好。另外，DeepSeek-V3 的论文犹如嘉年华大放送，详细介绍了从模型到基础设施的方方面面，并且还开源了，甚至讲解了生产级的 MoE 实现。 Polo Club 可视化：https://poloclub.github.io/transformer-explainer/ 本文也在我的微信公众号发表： Link: https://mp.weixin.qq.com/s/V_uyS-rOFVia3Mu49QhuHg PDF: LLM-Fundamentals-Visualizing-Transformer-Internals-CN.pdf 引用…
1. UB-Mesh: 基于统一互联和高维直连拓扑的AI集群架构 - 知返 https://zhuanlan.zhihu.com/p/32978184612 2. Bing 搜索核心技术 BitFunnel 原理 https://zhuanlan.zhihu.com/p/92840881?utm_id=0 1. Very good. This is generation technology breakthrough of search engine 1. text match -> BitFunnel 2. vector distance 2. BitFunnel: Revisiting Signatures for Search https://www.youtube.com/watch?v=1-Xoy5w5ydM 1. Higher rank rows 2. Frequency…
(Below are generated by AI translation. The original article was written in Chinese. See Vision and Strategy to the Storage Landscape (Chinese Simplified) .) Vision & Strategy: Insight, Foresight, and Strategy Vision and Strategy begin with asking questions: where should we be in 1 year, 3-5 years, or even 10 years? What should the team and departments be doing, and how should they be working?…
1. DeepSeek 3FS: https://github.com/deepseek-ai/3FS/tree/main 0. 3FS release highlights 1. Performance: 6TiB/s sustained throughput with 180 storage nodes => 300Gbps per node 2. Architecture: FoundationDB for cluster metadata, RocksDB for storage node metadata, CRAQ for consistent replication supported parallel reads 3. Rigor: both storage node and network (RDMA) protocol are formally modeled in…
1. A "Hitchhiker's" Guide to Fast and Efficient Data Reconstruction in Erasure-coded Data Centers [2014, SIGMOD14, 308 refs, Facebook] https://www.cs.cmu.edu/~nihars/publications/Hitchhiker_SIGCOMM14.pdf Slides: https://www.camdemy.com/media/16288 1. Related author "K. V. Rashmi". Apply Piggyback Framework on RS code, reduce both network by 25% and disk by 45%. Very good paper. 1. The key argument…
Recently, DeepSeek has emerged, not only reducing the cost of large language models by several times [1] , but also open-sourcing the entire AI infrastructure, the models, and the operational costs. Among them, 3FS demonstrates what an AI-oriented storage system looks like, deeply integrating and optimizing with DeepSeek’s AI infrastructure [4] . Many experts on the internet have already provided…
The speed of AI development is beyond imagination. For example, the VS Code Copilot Agent [1] can quickly read a large amount of code, understand the main components and interactions, and even draw class diagrams and flowcharts. It is truly a powerful tool for researching open-source projects. This article uses it to read the source code of DeepSeek 3FS [2] to generate explanations, riding the…
1. Convertible Codes: New Class of Codes for Efficient Conversion of Coded Data in Distributed Storage [2020, 22 refs, ITCS] https://www.pdl.cmu.edu/PDL-FTP/BigLearning/LIPIcs-ITCS-2020-66_abs.shtml 1. Convertible codes essentially work by merging multiple shorter codes into fewer longer codes. The 2020 paper focuses on MDS and "merge regime". Convertible codes have more follow up studies by this…
1. Run CUDA on other GPUs not from NVIDIA 1. Use CUDA without an NVIDIA GPU? [duplicate] https://stackoverflow.com/questions/68470042/use-cuda-without-an-nvidia-gpu 2. Can I use CUDA with a non-NVIDIA GPU? [duplicate] https://stackoverflow.com/questions/55320552/can-i-use-cuda-with-a-non-nvidia-gpu 3. 初创公司硬刚英伟达：在AMD卡上模拟CUDA，原版程序直接编译运行 https://zhuanlan.zhihu.com/p/709105635 4. HackerNews: Run CUDA,…
Vision & Strategy 洞察、远见与策略 Vision 与 Strategy 始于提问，1 年、3~5 年后、乃至 10 年后，我们应该处于什么位置，团队和部门应该在做什么，应该以什么样的方式工作。Vision 并不是指寻找最新技术趋势，并学习和应用它们。Vision 是“当事者”需要预测正确的技术趋势，确定投资方向，并用数据驱动和系统分析支持其的结论。 总的来说，其思考方式更接近于 产品经理 、 商业分析 （Business Analytics）、 市场调研 （Market Research），而不是技术开发工作。当然，向内，其也需要良好的技术基础（见 A Holistic View of Distributed Storage Architecture and Design Space…
Alibaba Cloud databases. PolarDB, OceanBase, AnalyticsDB. Huawei TaurusDB. AWS Aurora. 1. OceanBase Paetica: A Hybrid Shared-nothing/Shared-everything Database for Supporting Single Machine and Distributed Cluster [2023, 4 refs, VLDB23, Alibaba] https://www.vldb.org/pvldb/vol16/p3728-xu.pdf 1. Paetica is OceanBase 4.0. Maintain the performance of single node OceanBase comparable to single node…
1. Combining Buffered I/O and Direct I/O in Distributed File Systems [2024, 0 refs, FAST24] https://www.usenix.org/conference/fast24/presentation/qian https://www.youtube.com/watch?v=YPpTPQURy-U 1. Target HPC / Lustre. Dynamically switch Buffered IO and Direct IO. AutoIO, Simple but achieved 20% to 50% performance. improvement. First paper to try this approach. Not in paper but may apply similar…
1. AI时代云厂商仓皇的文件存储 - 王太平 太平说存储 [2024] https://mp.weixin.qq.com/s/Ir0bi5KYvPUfAw-qYZNM_g 1. Useful summary of filesystem landscape n. Related materials 1. GTC 2024 英伟达的存储观点 - 王太平 太平说存储 https://mp.weixin.qq.com/s/bCV1Efux_S8Ye7PunDlpqQ 0. Good. 1. Which storage to use? Local SSD => TOR NVMe => RDMA flash 2. GPUDirect Storage (GDS): Data path bypass CPU, partner ecosystem vendors, 3. Big accelerator…
Search “very good”, “good”, “very interesting”, “interesting”, “very useful”, “useful” for recommendation, and “my questions” for comments. 1. OpenAI Sora: Video generation models as world simulators https://openai.com/sora https://openai.com/research/video-generation-models-as-world-simulators 1. Sora text-to-video model is able to generate 1 minute of high resolution and high fidelity video…
1. 达利 原则 PRINCIPLES RAY DALIO 1. 前言 1. “正确地失败”是指，能够在经历痛苦的失败的过程中吸取重要的教训，从而避免“错误地失败”，即因为失败而被踢出局。 2. 共识性的观点通常已经反映在价格中了 2. 自传历程 1. 共识系统：股价反映了人们的预期，所以当实际结果比预期好时，股价上涨；当实际结果比预期差时，股价下跌。而大多数人会因为近期的经验而产生偏见。 2. 我之所以未能预见到这样的情况，是因为我对自我出生以来未曾发生过的而之前发生过很多次的事情感到惊奇。 3. 当所有人想法都一样时（例如都相信买入“漂亮50股”包赚不赔），这一情况几乎必然会反映在价格中，而把赌注押在这上面，就有可能犯错。 我把主要精力用来追踪市场动态，并设身处地地为我的企业客户展示，假如我是他们，我会如何处理市场风险。…
Search “very good”, “good”, “very interesting”, “interesting”, “very useful”, “useful” for recommendation, and “my questions” for comments. 1. 模型与算法在石油产业链的优化应用实践 - DataFunTalk https://zhuanlan.zhihu.com/p/680432343 2. 阿里云网络演进 - 木木女神经 https://zhuanlan.zhihu.com/p/680367683 3. 智算中心网络架构白皮书 - 木木女神经 https://zhuanlan.zhihu.com/p/680509531 4. LLM推理优化系统工程概述 - 进击的Killua…
Experience in cross BU communication and “Break Barriers and Build Consensus” A summary of skills I found - [Negotiation Skills] on non-easy discussions with stakeholders 1. First, before online, battle test your own conclusion and logic offline. Attack it with every direction. Think in-depth about every implication. Make notes. And group discussion with team internal first. 2. Be extremely…
1. QuePaxa: Escaping the Tyranny of Timeouts in Consensus [2023, SOSP23, 1 refs] https://expolab.org/ecs265-fall-2023/slices/QuePaxa-DDS.pdf https://bford.info/pub/os/quepaxa/quepaxa.pdf 1. Very good, paradigm shift work from the Paxos algorithm. First invention. 1. Paxos can be thought as each proposer is broadcasting itself to all replicas. In Paxos, each leader tries to shoot the others…
Misc articles. 1. AWS Fault Isolation Boundaries - AWS Whitepaper https://docs.aws.amazon.com/pdfs/whitepapers/latest/aws-fault-isolation-boundaries/aws-fault-isolation-boundaries.pdf#abstract-and-introduction 1. AZ, Region 2. AWS Local Zone - Edge location deployment 3. AWS Outpost - On-premises deployment 4. Static stability - interesting 1. AWS data planes statically-stable to an impairment in…
Picking my interested points. 1. AWS S3 Express One Zone 0. Overview of AWS re:Invent 2023 1. 芯片更大更强：服务器加速芯片Graviton4。AI推理训练芯片Trainum2。 2. AI: Amazon Q - ChatGPT。 Bedrock大模型平台。更多产品，逐渐全栈。 3. 云数据库：Aurora Limitless水平扩展，零RPO可用区切换，NVMe存储加速，Tiered缓存，Grover日志存储，Nitro芯片原子钟，Caspian热资源管理。 4. 存储：S3 Express One Zone。Zero-ETL集成。 1. What is S3 Express One Zone?…
The Entropy part. 1. PolarDB-X 存储引擎核心技术 | Lizard 分布式事务系统 https://zhuanlan.zhihu.com/p/656922735 https://zhuanlan.zhihu.com/p/654126910 1. interesting 2. 分布式事务模型 1. Percolator模型 2. Calvin模型 3. XA模型 3. 协调日志会被下沉到存储引擎上。其中一个参与方会被选为主分支。 2. ML system 入坑指南 - Fazzie https://zhuanlan.zhihu.com/p/608318764 1. useful materials and courses 3. 充分发挥NVMe存储的潜力：高性能存储引擎的设计与实现 - Andy730…
（下文由 AI 翻译生成，原文由英文写成，见 A Holistic View of Distributed Storage Architecture and Design Space 。） 本文总结软件架构方面的经验。架构设计本质上由 哲学 驱动，它是生产知识的引擎。从 组织视角 ，可以寻找架构设计的过程和技能为什么如此构建。通用的 方法论 和 原则 ，与哲学层面相联，为高质量的架构设计铺路。架构师需要针对不同 系统属性 （System Properties）的技术武器库。本文对分布式存储各个领域的 参考架构 （Reference Architecture）进行分类，从中总结 架构设计模式 。将它们互联，绘制 设计空间 （Design Space）的地图。 目录 软件架构-哲学视角 现实、语言和人类思维 关于软件架构 附：用例子解释 为什么需要软件架构 技术方面 抓住重要方面 过程与组织…
Large Language Model / LLM. Training in distributed framework and parallel models. GPT-4V. 8. ChatGPT 为什么不用 Reward-Model 的数据直接 fine-tune，而用 RL？ - Andy Yang https://www.zhihu.com/question/596230048/answer/3002469682 1. 多样性角度：不仅仅生成SL指定的“正确答案”, "rather than trying to replicate the human responses directly" 负反馈角度 自我知识感知角度：不要编造 n. related materials 1. Reinforcement Learning for Language Models…
The Entropy part. 1. 如何评价DuckDB? https://www.zhihu.com/question/438725169/answer/3143660836 1. "初衷是提供一个AP版的SQLite，即： 一个面向数据分析场景的嵌入式数据库" n. related materials 1. 万字解读 | A轮就融资￥3亿+的MotherDuck到底是个啥？ https://www.rachellaw.xyz/2023/MotherDuck 1. "MotherDuck的创始人Jordan是DuckDB的早期用户。 Jordan此前是Google BigQuery的创始工程师和产品经理" 2. MotherDuck: BIG DATA IS DEAD https://motherduck.com/blog/big-data-is-dead/ 1. "I…
Vector database and vector index. 1. VBASE: Unifying Online Vector Similarity Search and Relational Queries via Relaxed Monotonicity [2023, 0 refs, OSDI23] https://www.usenix.org/conference/osdi23/presentation/zhang-qianxi 1. Background: VectorDB becomes a fundamental infrastructure for Machine Learning and LLM. Evaluation: VBASE offers up to three orders-of-magnitude higher performance than…
Pangu 2.0 filesystem. 2. Pangu 2.0: More Than Capacity: Performance-oriented Evolution of Pangu in Alibaba [2023, 1 refs, FAST23, Alibaba] https://www.usenix.org/conference/fast23/presentation/li-qiang-deployed 1. Good paper. Comprehensive industry implementation for a cloud storage backend. Key point: Userspace filesystem USSOS driven by SSD and RDMA, leveraging DPDK and SPDK. Results: Figure 10,…
1. 学术研究中，有了一个不错的研究创意后该怎么办？ - 赵山 https://www.zhihu.com/question/315289489/answer/618549697 2. SQLite的文艺复兴 - Dawei Ma https://zhuanlan.zhihu.com/p/601510076 1. How the SQLite Virtual Machine Works - Ben Johnson https://fly.io/blog/sqlite-virtual-machine/ 3. Ray分布式计算框架详解 - liadrinz https://zhuanlan.zhihu.com/p/460600694 1. 字节跳动基于 Ray 的大规模离线推理 https://zhuanlan.zhihu.com/p/636971612 1. 大模型离线推理（Batch…
EC on SMR drives. 1. SMORE: A Cold Data Object Store for SMR Drives (Extended Version) [2017, 12 refs] https://arxiv.org/abs/1705.09701 1. Append-only. Cross SMR zone EC. Input data is stripped to fill zone. 2. GC needs to migrate live data out of zone. The EC pattern is essentially stripe/cut object to EC symbols. So no need to rewrite parity. 3. PMEM in front of HDD to coalescing writes (DRAM…
CockroachDB SSI, C++ coroutine, 文件/存储系统调研 1. 分布式文件系统调研（详细版）- 贺大伟 https://zhuanlan.zhihu.com/p/493647334 1. typical distributed filesystems 1. GFS 2. HDFS 3. Openstack Swift 4. MinIO 5. Ceph 6. TFS 7. Lustre 8. MooseFS 9. MogileFS 10. FastDFS 11. GlusterFS 12. GridFS 13. JuiceFS 14. ChubaoFS 15. Ozone 16. PolarFS 2. Cloudy with High Chance of DBMS: A 10-year Prediction for Enterprise-Grade ML - 约修亚…
Consultant thinking, analyzing business problems. 1. Readings: Vision: Case Interview Secrets - Victor Cheng https://www.caseinterview.com/ https://www.craftingcases.com/case-interview-examples/ 1. use proxy for estimation 1. match you intuition and find proxy via 1. relevance links 2. history or similar peers 3. population breakdown 2. improving the proxy 1. break the problem into different…
Search keywords for recommended papers: (as) (a) reference, reference (architecture). very good, good, interesting. very useful, useful. my question(s). Papers Custom Scheduling in Kubernetes, etc. 1. Compositional Model Checking of Consensus Protocols Specified in TLA+ via Interaction-Preserving Abstraction [2022, 1 refs, SRDS22 Best Paper] https://zhuanlan.zhihu.com/p/580744043…
Evaluate ChatGPT at different aspects in my daily work. 1. ChatGPT to read a paper and I ask questions 1. Try Facebook Tectonic https://chat.openai.com/chat/3c27228a-1e5e-4cf8-b2c5-9b9feda343fb 1. OK to answer basic questions 2. When I drill down to more detailed questions that need knowledge background, logic deduction, or deeper thinking .. ChatGPT starts to FAKE 2. Try "Improved Maximally…
Foundation papers of ChatGPT. 1. Transformer: Attention Is All You Need [2017, 69872 refs] https://arxiv.org/abs/1706.03762 1. Revolutionized Machine Translation area. Founding cornerstone for ChatGPT 2. highlights 1. key technologies - Prior work Seq2Seq model 1. Self-attention mechanism 1. No recurrence, no RNN. This allows fast parallel training with GPU 2. Multi-head attention 3. Positional…
The article summarizes my experiences on software architecture. Architecture design is essentially driven by philosophies as the generator engine that governs all knowledge. From the organization view , we can find why and how architecture design process and skills are required that way. Common methodologies and principles , viewed from the philosophies, provide guidance to carry out architecture…
As the new storage media being quickly adopted, it’s interesting how to manage cache on SSD device and work together with DRAM memory. 1. Readings: Papers: Managing Cache on SSD device 1. SSD as Cache: CacheLib and Kangaroo https://zhuanlan.zhihu.com/p/430451374 1. good, as introducing key designs of caching on SSD 2. papers 1. Facebook CacheLib 2. Kangaroo: Caching Billions of Tiny Objects on…
Networking is another pillar for distributed storage systems. 1. DCQCN: Congestion Control for Large-Scale RDMA Deployments [2015, 452 refs] https://conferences.sigcomm.org/sigcomm/2015/pdf/papers/p523.pdf 1. The recognized datacenter networking technology for RDMA RoCEv2, following up from the recognized DCTCP 2. highlights 1. problems of prior arts 1. Priority-based Flow Control (PFC) does not…
Archiving notes about recent reading articles. Many. 2. Misc: Readings: Recent piled up articles 1. Slack 架构设计 - Cocolian 老熊看架构 https://mp.weixin.qq.com/s/7RMYxu63juP4UHao0nMWkA 2. Whatsapp 系统架构 - Cocolian 老熊看架构 https://mp.weixin.qq.com/s/q1dE8xFRL5rTvfDITB6DYQ 3. Microsoft Dynamics 365 简史 - Cocolian 老熊看架构 https://mp.weixin.qq.com/s/Sw4U35EI6fkRKC47iuhcsg 微软PowerPlatform简史 - Cocolian 老熊看架构…
Tracking recent paper reading notes. For a better view, paste the notes into a text editor with larger line width. 3. Reading: Recent piled up papers 1. FAST'22 InfiniFS: An Efficient Metadata Service for Large-Scale Distributed Filesystems [2022, 0 refs, Alibaba] https://zhuanlan.zhihu.com/p/492210459 https://www.usenix.org/conference/fast22/presentation/lv 1. Good. Can be used as a Refernece…
Search tags to find recommended papers: “(can be used as a) reference (architecture)”, “very good”, “good”, “very interesting”, “interesting”, “very useful”, “useful” Papers about databases & storage 1. Casper: Optimal Column Layout for Hybrid Workloads (2020 VLDB, 17 refs) https://www.youtube.com/watch?v=AsjqfidHNAQ https://stratos.seas.harvard.edu/files/stratos/files/caspervldb2020.pdf 1. very…
Archiving notes about recent reading articles. Many. 2. Misc: Readings: Piled up Weichat articles (3hrs) 1. tidying up all opened tabs and move to bookmark and sync manually (2hrs) 1. WTF .. verbose .. hope chrome sync worked 2. 读 Instructions for PhD Students 有感 https://zhuanlan.zhihu.com/p/430479667 1. Instructions for PhD Students - Dimitris Papadias, HKUST…
Here to summarize my experience in how to do innovation in the daily engineering work. The focus area is distributed storage systems. Common Senses First, why engineering work needs innovation? Incremental innovation. Our daily work is to build better systems. The work is essentially composed of incremental innovation . Some proposed new features enroll into next level paradigm shift , or even…
Search good / (very) interesting / (very) useful, to find recommended papers. Search “question” to find feedback. Misc papers about Erasure Coding 1. Erasure Coding Stripping 1. RAID implements Erasure Coding by striping https://data-flair.training/blogs/hadoop-hdfs-erasure-coding/ 2. EC striped layout vs contiguous layout…
Search good / (very) interesting / (very) useful, to find recommended papers. Search “question” to find feedback. Industry papers. 1. Helios: Hyperscale Indexing for the Cloud & Edge (Microsoft) [2020, 2 refs] http://www.vldb.org/pvldb/vol13/p3231-potharaju.pdf 1. good. new direction to build global scale secondary index, eventual consistency, with log is the system/database principles. Hyperscale…
Query optimizer is the database signature technology, and among those most complex and less understood ones. Here I wrote down the key concepts, design paradigm, and how it works, from my recent study. Query Optimizer: Volcano/Cascades PDF WeiChat Related paper reading 1. Access Path Selection in Main-Memory Optimized Data Systems: Should I Scan or Should I Probe? [2017, 33 refs, CMU 15-721]…
Upon live production issue, before start, there are a few non-technical steps easy to neglect but proved very useful in practice Search the victim cluster name in mail or history, to beware of any issues other people already investigated or on-going, and any operations other people just done. Highly possible they will reveal hints to the current issue. The major sources of live production issues…
To design / evaluate / find optimal EC schema design, the general approach Understand the problem Looking for a faster EC schema? To reduce tail latency? To save capacity? To improve throughput? Better reliability? Etc Understand the scope and scenario: Cluster rack domain setup and placement constraints. Failure patterns, blast radius. Hardware replacement and software upgrading needs.…