1. Chord+DHash+Ivy: Building Principled Peer-to-Peer Systems Robert Morris rtm@lcs.mit.edu Joint work with F. Kaashoek, D. Karger, I. Stoica, H. Balakrishnan, F. Dabek, T. Gil, B. Chen, and A. Muthitacharoen

2. What is a P2P system? A distributed system architecture: No centralized control Nodes are symmetric in function Large number of unreliable nodes Enabled by technology improvements Node Internet

3. The promise of P2P computing High capacity through parallelism: Many disks Many network connections Many CPUs Reliability: Many replicas Geographic distribution Automatic configuration Useful in public and proprietary settings

4. Distributed hash table (DHT) Distributed hash table Distributed application get (key) data node …. put(key, data) Lookup service lookup(key)node IP address Application may be distributed over many nodes DHT distributes data storage over many nodes (Ivy) (DHash) (Chord)

5. A DHT has a good interface Put(key, value) and get(key)  value Call a key/value pair a “block” API supports a wide range of applications DHT imposes no structure/meaning on keys Key/value pairs are persistent and global Can store keys in other DHT values And thus build complex data structures

6. A DHT makes a good shared infrastructure Many applications can share one DHT service Much as applications share the Internet Eases deployment of new applications Pools resources from many participants Efficient due to statistical multiplexing Fault-tolerant due to geographic distribution

7. Many recent DHT-based projects File sharing [CFS, OceanStore, PAST, …] Web cache [Squirrel,..] Backup store [Pastiche] Censor-resistant stores [Eternity, FreeNet,..] DB query and indexing [Hellerstein, …] Event notification [Scribe] Naming systems [ChordDNS, Twine,..] Communication primitives [I3, …] Common thread: data is location-independent

8. The lookup problem Internet N1N1 N2N2 N3N3 N6N6 N5N5 N4N4 Publisher Put (Key=“title” Value=file data…) Client Get(key=“title”) ? At the heart of all DHTs

9. Centralized lookup (Napster) Publisher@ Client Lookup(“title”) N6N6 N9N9 N7N7 DB N8N8 N3N3 N2N2 N1N1 SetLoc(“title”, N4) Simple, but O( N ) state and a single point of failure Key=“title” Value=file data… N4N4

10. Flooded queries (Gnutella) N4N4 Publisher@ Client N6N6 N9N9 N7N7 N8N8 N3N3 N2N2 N1N1 Robust, but worst case O( N ) messages per lookup Key=“title” Value=file data… Lookup(“title”)

11. Routed queries (Freenet, Chord, etc.) N4N4 Publisher Client N6N6 N9N9 N7N7 N8N8 N3N3 N2N2 N1N1 Lookup(“title”) Key=“title” Value=file data…

12. Chord lookup algorithm properties Interface: lookup(key)  IP address Efficient: O(log N) messages per lookup N is the total number of servers Scalable: O(log N) state per node Robust: survives massive failures Simple to analyze

13. Chord IDs Key identifier = SHA-1(key) Node identifier = SHA-1(IP address) SHA-1 distributes both uniformly How to map key IDs to node IDs?

14. Chord Hashes a Key to its Successor N32 N10 N100 N80 N60 Circular ID Space Successor: node with next highest ID K33, K40, K52 K11, K30 K5, K10 K65, K70 K100 Key ID Node ID

15. Basic Lookup N32 N10 N5 N20 N110 N99 N80 N60 N40 “Where is key 50?” “Key 50 is At N60” Lookups find the ID’s predecessor Correct if successors are correct

16. Successor Lists Ensure Robust Lookup N32 N10 N5 N20 N110 N99 N80 N60 Each node remembers r successors Lookup can skip over dead nodes to find blocks N40 10, 20, 32 20, 32, 40 32, 40, 60 40, 60, 80 60, 80, 99 80, 99, 110 99, 110, 5 110, 5, 10 5, 10, 20

17. Chord “Finger Table” Accelerates Lookups N80 ½ ¼ 1/8 1/16 1/32 1/64 1/128

18. Chord lookups take O(log N) hops N32 N10 N5 N20 N110 N99 N80 N60 Lookup(K19) K19

19. Simulation Results: ½ log 2 (N) Number of Nodes Average Messages per Lookup Error bars mark 1 st and 99 th percentiles

20. DHash Properties Builds key/value storage on Chord Replicates blocks for availability Caches blocks for load balance Authenticates block contents

21. DHash Replicates blocks at r successors N40 N10 N5 N20 N110 N99 N80 N60 N50 Block 17 N68 Replicas are easy to find if successor fails Hashed node IDs ensure independent failure

22. DHash Data Authentication Two types of DHash blocks: Content-hash: key = SHA-1(data) Public-key: key is a public key, data is signed by that key DHash servers verify before accepting Clients verify result of get(key)

23. Ivy File System Properties Traditional file-system interface (almost) Read/write for multiple users No central components Trusted service from untrusted components

24. Straw Man: Shared Structure Standard meta-data in DHT blocks? What about locking during updates? Requires 100% trust Root Inode Directory Block File3 Inode File2 Inode File1 Inode File3 Data

25. Ivy Design Overview Log structured Avoids in-place updates Each participant writes only its own log Avoids concurrent updates to DHT data Each participant reads all logs Private snapshots for speed

26. Ivy Software Structure App NFS Client Ivy Server Internet DHT Node DHT Node DHT Node kernel user system calls NFS RPCs

27. One Participant’s Ivy Log Log Head Record 1 Record 2 Record 3 Mutable public-key signed DHT block Immutable content-hash DHT blocks Log-head contains DHT key of most recent record Each record contains DHT key of previous record

28. Ivy I-Numbers Every file has a unique I-Number Log records contain I-Numbers Ivy returns I-Numbers to NFS client NFS requests contain I-Numbers In the NFS file handle

29. NFS/Ivy Communication Example Local NFS ClientLocal Ivy Server LOOKUP(“d”, I-Num=1000) I-Num=1000 CREATE(“aaa”, I-Num=1000) I-Num=9956 WRITE(“hello”, 0, I-Num=9956) OK echo hello > d/aaa LOOKUP finds the I-Number of directory “d” CREATE creates file “aaa” in directory “d” WRITE writes “hello” at offset 0 in file “aaa”

30. Log Records for File Creation Type: Create I-num: 9956 Type: Link Dir I-num: 1000 File I-num: 9956 Name: “aaa” Type: Write I-num: 9956 Offset: 0 Data: “hello” … Log Head A log record describes a change to the file system

31. Scanning an Ivy Log Type: Link Dir I-num: 1000 File I-num: 9956 Name: “aaa” Type: Link Dir I-num: 1000 File I-num: 9876 Name: “bbb” Type: Remove Dir I-num: 1000 Name: “aaa” A scan follows the log backwards in time LOOKUP(name, dir I-num): last Link, but stop at Remove READDIR(dir I-num): accumulate Links, minus Removes

32. Finding Other Logs: The View Block Log Head 1 Log Head 2 Log Head 3 View block is immutable (content-hash DHT block) View block’s DHT key names the file system Example: /ivy/37ae5ff901/aaa View Block Pub Key 1 Pub Key 2 Pub Key 3

33. Reading Multiple Logs Log Head 1 Log Head 2 Problem: how to interleave log records? Red numbers indicate real time of record creation But we cannot count on synchronized clocks 20 273132 3029282726

34. Vector Timestamps Encode Partial Order Log Head 1 Log Head 2 20 273132 3029282726 Each log record contains vector of DHT keys One vector entry per log Entry points to log’s most recent record

35. Snapshots Scanning the logs is slow Each participant keeps a private snapshot Log pointers as of snapshot creation time Table of all I-nodes Each file’s attributes and contents Reflects all participants’ logs Participant updates periodically from logs All snapshots share storage in the DHT

36. Simultaneous Updates Ordinary file servers serialize all updates Ivy does not Most cases are not a problem: Simultaneous writes to the same file Simultaneous creation of different files in same directory Problem case: Unlink(“a”) and rename(“a”, “b”) at same time Ivy correctly lets only one take effect But it may return “success” status for both

37. Integrity Can attacker corrupt my files? Not unless attacker is in my Ivy view What if a participant goes bad? Others can ignore participant’s whole log Ignore entries after some date Ignore just harmful records

38. Ivy Performance Half as fast as NFS on LAN and WAN Scalable w/ # of participants These results were taken yesterday…

39. Local Benchmark Configuration App NFS Client Ivy Server DHash Server One log One DHash server Ivy+DHash all on one host

40. Ivy Local Performance on MAB PhaseIvyNFS Mkdir0.60.5 Write6.60.8 Stat0.60.2 Read1.00.8 Compile10.05.3 Total18.87.6 Modified Andrew Benchmark times (seconds) NFS: client – LAN – server 7 seconds doing public key signatures, 3 in DHash

41. WAN Benchmark Details 4 DHash nodes at MIT, CMU, NYU, Cornell Round-trip times: 8, 14, 22 milliseconds No DHash replication 4 logs One active writer at MIT Whole-file read on open() Whole-file write on close() NFS client/server round-trip time is 14 ms

42. Ivy WAN Performance PhaseIvyNFS Mkdir3.21.8 Write24.69.6 Stat12.09.6 Read11.412.0 Compile41.427.8 Total92.660.8 47 seconds fetching log heads, 4 writing log head, 16 inserting log records, 22 in crypto and CPU

43. Ivy Performance w/ Many Logs MAB on 4-node WAN One active writer Increasing cost due to growing vector timestamps

44. Related Work DHTs Pastry, CAN, Tapestry File systems LFS, Zebra, xFS Byzantine agreement BFS, OceanStore, Farsite

45. Summary Exploring use of DHTs as a building block Put/get API is general Provides availability, authentication Harnesses decentralized peer-to-peer groups Case study of DHT use: Ivy Read/write peer-to-peer file system Trustable system from untrusted pieces http://pdos.lcs.mit.edu/chord