1. Coordination in Distributed Systems Lecture # 8

2. Coordination Anecdotes  Decentralized, no coordination  Aloha ~ 18%  Some coordinating Master  Slotted Aloha ~ 36%  Impossibility of Concensus  If the ether dropped your packets, we will never have concensus (lynch)  Decentralized, no coordination  Aloha ~ 18%  Some coordinating Master  Slotted Aloha ~ 36%  Impossibility of Concensus  If the ether dropped your packets, we will never have concensus (lynch)

3. Why is coordination important  Resource Sharing  Database  Files  Agreement  Agree on an award for faculty  Agree on whether to take a quiz  Leader Election  Choose Time server  Choose Bus Master  Resource Sharing  Database  Files  Agreement  Agree on an award for faculty  Agree on whether to take a quiz  Leader Election  Choose Time server  Choose Bus Master

4. Challenges  Host Failures  Link Failures  Time and Causual Ordering  Host Failures  Link Failures  Time and Causual Ordering

5. Distributed Mutual Exclusion  Mutual Exclusion  Atmost one process in the critical section  What happens when processes are distributed  Let’s assume reliability first  Three messages  Enter(): I want to enter  ResourceAccess(): OK Go ahead access the critical section  Exit(): I am done  Mutual Exclusion  Atmost one process in the critical section  What happens when processes are distributed  Let’s assume reliability first  Three messages  Enter(): I want to enter  ResourceAccess(): OK Go ahead access the critical section  Exit(): I am done

6. Requirements for Mutex  At most one process in the critical section  Everyone gets to access the critical section: Eventual grant  Global ordering: Difficult  At most one process in the critical section  Everyone gets to access the critical section: Eventual grant  Global ordering: Difficult

7. Distributed Mutual Exclusion P1 P2 P3 P4

8. Centralized Algorithm  Single Coordinating Server  P can enter Critical Section (CS) if it has a grant TOKEN  Server Grants TOKEN  Processes Request TOKEN  GRANT if CS is free  Else queue  Grant to the oldest in the queue  Problems  Performance Bottleneck  Server Crash  Ordering?  Single Coordinating Server  P can enter Critical Section (CS) if it has a grant TOKEN  Server Grants TOKEN  Processes Request TOKEN  GRANT if CS is free  Else queue  Grant to the oldest in the queue  Problems  Performance Bottleneck  Server Crash  Ordering?

9. Ring Based Coordination Logical Ring Typically based on network address Pass token clockwise (other otherwise)

10. Ring based Coorindation  No Master, no Centralized Server  Wait for TOKEN  If you want to enter CS  Grab Token  Access CS  Release and Pass along the TOKEN  Else  Pass along the Token  Problems:  Continuous Overhead  Must wait for N (worst case) before enter is granted  Ordering?  No Master, no Centralized Server  Wait for TOKEN  If you want to enter CS  Grab Token  Access CS  Release and Pass along the TOKEN  Else  Pass along the Token  Problems:  Continuous Overhead  Must wait for N (worst case) before enter is granted  Ordering?

11. Ricarta&Agarwal Algo  Based on Reliable Multicast  Multicast to All Processes when you want to enter CS  If you hear an Ack from everyone  Enter CS  Else  Repeat  Ordering?  Timestamp multicast with Lamport Clocks  Use vector clocks for total ordering  Based on Reliable Multicast  Multicast to All Processes when you want to enter CS  If you hear an Ack from everyone  Enter CS  Else  Repeat  Ordering?  Timestamp multicast with Lamport Clocks  Use vector clocks for total ordering

12. Leader Election  Elect Unique Master  Typically after failure  Simulataneous Requests Allowed  Permissible States  Either No Leader  Or Well-known leader  Everyone must know about leader  All processes must participate in election and can discover leader  Elect Unique Master  Typically after failure  Simulataneous Requests Allowed  Permissible States  Either No Leader  Or Well-known leader  Everyone must know about leader  All processes must participate in election and can discover leader

13. Chang&Roberts Algo

14. Ring coordination  Find the Master with the maximum UID  Receive UID  Pass max(Own UID, Nieghbor UID)  If you hear your UID again  You are the leader  Annouce yourself as the leader  Problems:  Two Iterations  Unique IDs assumed  Failure?  Find the Master with the maximum UID  Receive UID  Pass max(Own UID, Nieghbor UID)  If you hear your UID again  You are the leader  Annouce yourself as the leader  Problems:  Two Iterations  Unique IDs assumed  Failure?

15. Itai&Rodeh Algo  Relaxes UID Requirement  Again Ring Based (unidirectional)  Election  Assign you’re an ID (from 1 … k)  Do the Chang algo  If no leader  Choose ID again  Termination?  Probablistic  Relaxes UID Requirement  Again Ring Based (unidirectional)  Election  Assign you’re an ID (from 1 … k)  Do the Chang algo  If no leader  Choose ID again  Termination?  Probablistic

16. Termination  Larger the K, larger your probability of assigning yourself a unique ID  N = 4, K=16  Time to terminate: 1.01 rounds  Welcome to the power of randomization  Used in everyday things to make problems tractable  Think Ethernet  Larger the K, larger your probability of assigning yourself a unique ID  N = 4, K=16  Time to terminate: 1.01 rounds  Welcome to the power of randomization  Used in everyday things to make problems tractable  Think Ethernet

17. IEEE 1394: Firewire  High-speed synchronous Bus  Used on Sony Handycams  Bus Master Election Algorithm  Not a ring (actually a tree)  Hot Plug-n-Play: NO UIDs  Must elect master after a device is attached or removed  High-speed synchronous Bus  Used on Sony Handycams  Bus Master Election Algorithm  Not a ring (actually a tree)  Hot Plug-n-Play: NO UIDs  Must elect master after a device is attached or removed

18. Algo  Start with Leaves  Send “adopt me”  Leave has only one egress link  If you hear n-1 “adopt me” messages  Forward the message to the Nth link  Why is probability used again for root detection?  Start with Leaves  Send “adopt me”  Leave has only one egress link  If you hear n-1 “adopt me” messages  Forward the message to the Nth link  Why is probability used again for root detection?