传统画流程图的痛点 我们经常需要画流程图来表示代码逻辑或者基本框架等。但我们在绘画流程图的时候，经常会在对齐连接线这些和流程图表达的意义无关的

beamer和powerpoint的不同 我们经常需要用ppt这一形式来展示我们的工作成果，但众所周知，微软的powerpoint是收费软件，

Ceph Storage: The Storage Powerhouse in the Era of AI/ML Workloads Abstract AI/ML training, inference, and related processes place unprecedented demands on storage performance. This article, based on the SNIA presentation “Ceph Storage in a World of AI/ML Workloads”, analyzes the challenges of AI storage, the advantages of Ceph, and key methods to improve efficiency in real deployments. AI/ML Workload Lifecycle A typical AI/ML lifecycle includes: Raw Data → Training Data → Model → Results → Retraining During training, network bandwidth, data preprocessing capability, and model size all affect overall performance.

This document provides a detailed analysis of the major feature evolution in Ceph from Nautilus (v14) to the latest Squid (v19) versions, offering guidance for selecting appropriate versions and developing upgrade strategies. (Organized with LLM assistance) Version Overview Version Codename Release Date Lifecycle Status v14.2.x Nautilus 2019 EOL v15.2.x Octopus 2020 EOL v16.2.x Pacific 2021 EOL v17.2.x Quincy 2022 EOL v18.2.x Reef 2023 Stable Maintenance v19.2.x Squid 2024 Current Stable Version Feature Comparison Summary Maturity-Driven Version Selection Guide Feature Nautilus Octopus Pacific Quincy Reef Squid Deployment Method ceph-deploy cephadm introduced cephadm cephadm mature cephadm cephadm Storage Engine BlueStore BlueStore BlueStore BlueStore FileStore removed BlueStore optimized Configuration Management Centralized introduced Centralized Centralized Centralized Centralized Centralized Network Protocol msgr2 introduced msgr2 stable msgr2 msgr2 msgr2 msgr2 PG Management autoscale introduced autoscale autoscale autoscale autoscale autoscale Scheduler Traditional Improved mclock introduced mclock default mclock mclock optimized CephFS Multi-FS First support Feature enhanced Mirroring perfected Management optimized Management optimized Dashboard integrated Multi-site Basic RBD mirroring CephFS mirroring Perfected Enhanced Enhanced Dashboard Basic Improved Improved Improved Refactored Refactored Containerization None cephadm preview cephadm mature cephadm complete cephadm cephadm Feature Maturity Marking Legend Bold text: Important milestones for features in this version (first introduction/stability achieved/major improvements) Normal text: Features remain stable or have minor improvements in this version Italic text: Features are deprecated or being prepared for removal in this version Feature-Driven Version Selection Guide Required Feature Minimum Version Stable Recommended Version Notes Centralized Configuration Management Nautilus Octopus+ Basic functionality available, upgrade recommended for stability PG autoscaling Nautilus Pacific+ Production environments recommend manual control msgr2 Security Protocol Nautilus Octopus+ Recommended for new deployments cephadm Container Management Octopus Pacific+ Tech preview → production ready CephFS Multi-filesystem Nautilus Pacific+ Basic support → production ready CephFS Mirroring/DR Octopus Pacific+ Feature introduction → production stable mclock QoS Scheduling Pacific Quincy+ Introduction → default enabled Full Containerized Deployment Pacific Quincy+ Basic support → complete ecosystem Advanced Dashboard Reef Squid+ Refactored → complete FileStore Replacement Any version Reef+ FileStore support removed after Reef Feature First Introduction and Maturity Analysis Important Feature Lifecycle Timeline This section details the first introduction version, stable version, and recommended production environment adoption timing for key features, helping users make informed version choices.

摘要 AI/ML 训练、推断等环节对存储提出了前所未有的高性能要求。本文结合 SNIA《Ceph Storage in a World of AI/ML Workloads》演示文稿内容，分析了 AI 存储