The Chubby lock service for loosely- coupled distributed systems
3 likes4,317 views
The document describes Chubby, a lock service developed by Google for loosely coupled distributed systems. Chubby uses a consensus protocol to elect a master from replicas and provides coarse-grained locking. It interfaces like a file system but stores ephemeral and permanent nodes. Clients cache data and use handles to subscribe to events. Chubby scales through techniques like reducing communication and partitioning while its usage includes locking, naming, and master election in distributed systems like GFS and Bigtable.
The Chubby lock service for loosely- coupled distributed systems
- 1. Advanced Topics in Distributed Systems The Chubby lock service for loosely- coupled distributed systems Mike Burrows, Google Inc. Presented by: Ioanna Tsalouchidou 1
- 2. Outline ● Introduction ● Design ● Scaling ● Usage ● Summary 2
- 3. Introduction ● What is Chubby? – An engineering effort by Google – Coarse-grained lock service – Loosely-coupled distributed systems – Large number of small machines – High-speed network (1Gbit/s Ethernet) – Availability and Reliability Vs. Performance 3
- 4. Outline ● Introduction - Motivation ● Design ● Scaling ● Usage ● Summary 4
- 5. Design le iona Rat ks Se Loc qu en Sy st em ce str rs uc tur rs e handle File Fa Backup es il-o pAliv ve Kee rs Dir Mirroring ec tor Events ies ns Sessio Ca se ch a Ba API ing Dat 5
- 6. Rationale ● Lock service library Vs. Paxos client library – Maintain existing program structure – Advertising the results – Familiar interface to programmers – Reduces the # of servers needed for progress ● Design decision – Small-files service (for primaries) – Event notification – Consistent caching 6
- 7. Rationale ● Coarse-grained locks – Less load in the lock server – Less delay caused by lock server unavailability – Survive lock server failures – Many clients served by less servers ● Fine-grained locks – Lock-server unavailability → many clients stall 7
- 8. System Structure ● One Chubby cell → five servers ● Client ↔ Server: via client library ● Server and library communicate via RPC 8 ● Optional: proxy server
- 9. Chubby Cell ● Set of five servers (replicas) :-) ● Distributed consensus protocol ● One master among replicas ● Master lease → periodically renewed 9
- 10. Chubby Client ● Clients find the master between replicas ● Client requests to master ● Write requests via consensus protocol among replicas ● Read requests handled by master App Library 10
- 11. Files, directories and handles ● Interface: similar to UNIX file system /ls/foo/wombat/pouch 11
- 12. Files, directories and handles ● Interface: similar to UNIX file system /ls/foo/wombat/pouch Stands for lock service. Common to all Chubby names 12
- 13. Files, directories and handles ● Interface: similar to UNIX file system /ls/foo/wombat/pouch Stands for lock service. Common to all Chubby names The name of a Chubby cell, which is resolved to one or more servers through a DNS lookup 13
- 14. Files, directories and handles ● Interface: similar to UNIX file system /ls/foo/wombat/pouch Stands for lock service. Common to all Chubby names The name of a Chubby cell, which is resolved to one or more servers through a DNS lookup The name of the directory and the file 14
- 15. Files, directories and handles ● Interface: similar to UNIX file system /ls/foo/wombat/pouch ● Name space: files and directories called nodes – Ephemeral – Permanent ● Clients open nodes to obtain Handles – ~ UNIX file descriptors – Subscription to events 15
- 16. Locks ● Any node can act as a lock ● Advisory locks (no mandatory locks) ● Client handles hold – Exclusive locks (write) – Shared locks (read) ● No write permission → no influence to writers 16
- 17. Sequencers ● Sequence numbers to interactions that use locks ● Lock holder requests a sequencer ● Opaque bite-string → state of lock – Name of lock – Mode which was acquired – Lock generation number 17
- 18. Caching ● To reduce read traffic – File data – Meta-data – File absence – Consistent, write-through cache 18
- 19. Fail-overs 19
- 20. Fail-overs 20
- 21. Fail-overs OOOPS 21
- 22. Fail-overs OOOPS 22
- 23. Outline ● Introduction - Motivation ● Design ● Scaling ● Usage ● Summary 24
- 24. Current scaling techniques ● Reduce communication with the master – Client caching – Open handles ● Use a nearby cell (found with DNS) ● Increase lease time (12s → 60s) ● Protocol-conversion servers – Translate Chubby to less-complex 25
- 25. Future scaling techniques ● Proxies – Handling KeepAlive requests – Read requests ● Partitioning – Enable large Chubby cells – Little communication between partitions 26
- 26. Outline ● Introduction ● Design ● Scaling ● Usage ● Summary 27
- 27. Usage ● Initially as a lock service ● Most popular use as a name service ● GFS: to appoint a master server ● Bigtable – Master election – Master discovers its controlling servers – Clients to find their master 28
- 28. Summary ● Distributed lock service – Coarse grained synchronization ● Based on existing ideas and algorithms ● Primary internal name service ● Reliable but low-volume storage 29
- 29. Advanced Topics in Distributed Systems The Chubby lock service for loosely- coupled distributed systems Mike Burrows, Google Inc. Presented by: Ιωάννα Τσαλουχίδου 30