CloudLab: A File System Friendly Key Value Store
5 likes1,163 views
The document discusses MaxiScale, a software platform that provides a file-system friendly key-value store for cloud-scale file serving and storage. It addresses the challenges of storing and serving billions of small objects cost-effectively. MaxiScale delivers horizontal scaling, high performance, availability and efficiency using commodity hardware through its distributed architecture and optimization of small file operations. Case studies demonstrate how MaxiScale improves performance and reduces costs for social networking and photo serving applications handling large volumes of user-generated content.
![User Profiles Use Key Value Store in SFR to store user profiles Single I/O file retrieval Data record up to 1 MB [email_address] <id>joe@maxiscale.com <password>***** <album set>vacations <album>florida <images> IMG_01234.JPG, IMG_04567.JPG, IMG_06789.JPG <thumbnail bundle 96 pixels> vacations.florida.thumbs96 <album set>family <album>kid Page April 2010 MaxiScale](https://image.slidesharecdn.com/cloudlabkv02-100427143030-phpapp02/85/CloudLab-A-File-System-Friendly-Key-Value-Store-26-320.jpg)
CloudLab: A File System Friendly Key Value Store
- 1. A File-System Friendly Key Value Store Cloud-Scale File Serving and Storage MaxiScale FLEX Software Platform @garyorenstein | go@maxiscale.com April 2010 Page MaxiScale
- 2. The Era of Billions 10 billion Tweets 10 billion downloads 3 billion photos/month 12 billion views January April 2010 Page MaxiScale
- 3. Beyond Capacity to Users and Objects Everyone is a content consumer AND creator April 2010 Page MaxiScale
- 4. Impact of The Era of Billions Companies face relentless pressure to store and serve data cost-effectively April 2010 Page MaxiScale
- 5. Cloud Applications Requirements: Fast, Scalable, Affordable “Light” Structured Data Retrieval Page April 2010 MaxiScale
- 6. Basics Anatomy of a Key-Value Store April 2010 Page Source: http://dominicfallows.co.uk MaxiScale
- 7. MaxiScale Key Value Store Benefits Page April 2010 MaxiScale
- 8. Breakthrough FLEX Software Platform Software Infrastructure On commodity hardware Targeted to Web companies Service Providers Enterprises Delivers file serving and storage to web and scale-out applications Web / App Servers File Serving and Storage April 2010 Page Internet MaxiScale
- 9. MaxiScale FLEX Platform: Core Technologies Distributed Platform Single Namespace Distributed Metadata and Data Peer Sets Optimized File Serving Unified Management Distributed MapReduce-enabled Planned: Custom apps (Hadoop) Linux and Windows Clients Small File Performance (key-value store) High-Density Storage Scale-Out NAS April 2010 Page MaxiScale Storage I/O Bandwidth CPU
- 10. Simplicity of a Single Namespace April 2010 Page Small File Performance (key-value store) Scale-Out NAS Single mount point / /NFR/ /SFR/ Directories Large files Small objects Flat hierarchy Screaming performance MaxiScale
- 11. FLEX Software Platform In Action Utmost system efficiency from client to network to cluster Peer Set April 2010 Page MaxiScale
- 12. Fast Lookups Directly from Disk Return values up to 1MB in a single disk I/O operation Automatically create file and index entry using POSIX open() call Read index, then metadata and data in single disk I/O Index maps to disk blocks for single disk I/O file retrieval April 2010 Page Index (RAM) Metadata and Data (Disk) MaxiScale
- 13. Performance Counts Customer results 4 nodes 80KB files 736 files/second 63.6 million per day Files/spindle/sec Read tests with 32KB files and small file optimizations April 2010 Page MaxiScale
- 14. Selecting a Key Value Store April 2010 Page MaxiScale
- 15. Accelerating Web content serving Page Brute-force speed-up Hardware and development cost Performance Efficient key-value lookup without database queries Slash cost Fast lookup Web-scale April 2010 MaxiScale Do-it-yourself key value store RAM Cache RAM Cache Relational DB (MySQL …) Relational DB (MySQL …) MaxiScale key value store
- 16. Page Which key value store and file system are right for your application? High Availability? Key Length? Lookup Latency? Number of Keys? Cost? Size of Values? Data Storage Density? April 2010 MaxiScale
- 17. Key value store landscape Page Circles represent relative usable capacity at load Memcache High RAM cost in large quantities Cost MaxiScale Single disk I/O Reads Petabyte scale Relational DB Long lookup latency File systems Long read latency Solid state disk High flash cost in large quantities Billions of keys; auto- expandable Key length < 1000 characters Lookup latency 5-10ms per key Stored values up to 1MB Highly available storage Latency April 2010 MaxiScale
- 18. Placing Key Value Stores in Context April 2010 Page MaxiScale Relational Databases Key-Value Stores File System Friendly Key-Value Stores File Systems SQL Joins ACID Big Data Billions of Small Objects Application Flexibility (unique data access mechanism) Big Data Billions of Small Objects Filename as key (or system generated key) POSIX compliance Same namespace as file system Large Objects Petabytes of content Directory structures POSIX compliance
- 19. Social Networking Case Study Page April 2010 MaxiScale
- 20. Social networking application example Challenges Expensive filer hardware 1000’s of mount points to manage Difficult to scale at 30% year-over-year growth Before MaxiScale April 2010 Page 175 TB content repository across multiple NAS systems Sharded database instances for profiles, messages… Application Servers Web Servers MaxiScale
- 21. Social networking application example Benefits Single storage mount point per server Serve photos and video clips in 5 milliseconds With MaxiScale April 2010 Page Application Servers Web Servers Financial savings Capital expense: 60% Operating expense: $105K First-year ROI: 187% News feed, message, and profile data Direct photo retrieval and upload MaxiScale Value
- 22. A Deeper Look at Photo Serving Case Study Page April 2010 MaxiScale
- 23. Examples of Photo Serving Requirements Page April 2010 MaxiScale
- 24. Originals in NFR POSIX-compliant hierarchy /nfr/joe@maxiscale.com /vacations /florida High-Resolution Original Images IMG_01234.JPG IMG_04567.JPG IMG_06789.JPG Page April 2010 MaxiScale
- 25. Generate thumbnails /vacations.florida. thumb96 .img01234.jpg /vacations.florida. thumb480 .img01234.jpg /vacations.florida. thumb640 .img01234.jpg Pre-Rendered Thumbnails IMG_01234.JPG 96 pixels wide 480 pixels wide 640 pixels wide Application Process Page April 2010 MaxiScale
- 26. User Profiles Use Key Value Store in SFR to store user profiles Single I/O file retrieval Data record up to 1 MB [email_address] <id>joe@maxiscale.com <password>***** <album set>vacations <album>florida <images> IMG_01234.JPG, IMG_04567.JPG, IMG_06789.JPG <thumbnail bundle 96 pixels> vacations.florida.thumbs96 <album set>family <album>kid Page April 2010 MaxiScale
- 27. MaxiScale Photo Serving Features and Benefits Page April 2010 MaxiScale MaxiScale Feature Benefit POSIX-compliant file system hierarchy with NFR Segment users, album sets, and albums into traditional directory structure Small File Repository (SFR) Single I/O file retrieval Thumbnails Profiles Key Value Store Single I/O profile retrieval Billions of keys Up to 1 MB value/record length Photo serving feature set leads to reduced infrastructure costs to meet user demand.
- 28. Delivering Cloud-Scale File Serving and Storage New scale-out application requirements Horizontal scale, linear performance Full availability and resiliency Utmost cost-efficiency MaxiScale delivers Cloud-Scale file serving and storage software Simple and always online Game changing cost using commodity hardware File System Friendly Key Value Store Page MaxiScale helps companies handle relentless pressure to store and serve data cost-effectively April 2010 MaxiScale
- 29. Thank You @garyorenstein | go@maxiscale.com @maxiscale | www.MaxiScale.com Page April 2010 MaxiScale