Elections in a Distributed Computing System

1. Elections in a Distributed Computing System Hector Garcia-Molina Presenter: Srinath Rao

2. Introduction to Elections • Strategies to deal with a node failure • Have s/w which can operate continuously even as failures occur • Halt temporarily, reorganize the system • Need for the coordinator and hence election • Election protocols can be used to start up a system, add/remove nodes • Issues ?

3. Issues • Constituent nodes may fail after election • What does it mean to be a coordinator? • How to cope up with failures during the election itself? • Always possible to select a unique coordinator? • Might wish to have more than one coordinator

4. Outline • Assumptions • Elections with no commn. failures • The Bully Election Algorithm • Elections with commn. failures • The Invitation Election Algorithm • Related Work • Conclusions

5. Assumptions • All nodes cooperate • Election algorithm makes use of “bug-free” software facilities • Communication subsystem will not spontaneously generate messages • Nodes have “safe” storage cells • Node halts processing when it fails • No transmission errors

6. Assumptions (contd..) • Messages are processed in the order they are received • Communication subsystem does not fail • Node never pauses

7. State Vector of node • A collection of safe storage cells • Principal components of Vector S(i) • Status of node i: S(i).s • Down, Election, Reorganization, Normal • Coordinator according to node i: S(i).c • Definition of the Task being performed: S(i).d

8. Outline • Assumptions • Elections with no commn. failures • The Bully Election Algorithm • Elections with commn. failures • The Invitation Election Algorithm • Related Work • Conclusions

9. Desired Characteristics • Assertion 1: For two nodes i and j • S(i).c = S(j).c if nodes i and j are in one of the states “Normal” or “Reorganization” • S(i).d = S(j).d if both i and j are in “Normal” state • States what it means to be a coordinator

10. Desired Characteristics (contd..) • Assertion 2: If no failures, election will eventually transform a system in any state to a state: • There exists node i with S(i).s =“Normal” and S(i).c = i • Other active nodes j have S(j).s = “Normal” and S(j).c = i

11. The Bully Election Algorithm • Each node has an unique id no. • Algorithm uses id no. as priorities • Two step algorithm • Node i tries to contact all nodes with higher priorities. If no reply received, then assume the role of coordinator • Inform all the lower priority nodes • Send “halt” message, force state of j to “Election” • Send “I am elected” message, node j sets S(j).c=I and S(j).s = “Reorganization” • Distribute new algorithms to nodes, all status changed to “Normal”

12. Bully (contd..) • Let the recovering node k attempt to become the coordinator using the same algorithm • Halts all lower priority nodes which may be in the process of becoming coordinators • Step 1 ensures no conflict with higher priority nodes

13. Outline • Assumptions • Elections with no commn. failures • The Bully Election Algorithm • Elections with commn. failures • The Invitation Election Algorithm • Related Work • Conclusions

14. Discussion • Failures • Partitioning of nodes • A node can only send/receive message • Node i and node j can talk to node k but not with each other • Node may pause and then resume • Observation: Impossibility of consensus in the event of failure of commn. subsystem or node pausing. • Inference: Redefine the meaning of an election

15. Discussion (contd..) • Notion of a group of nodes and group id • Node i stores group id in its state vector: S(i).g • Nodes are free to change groups • Identify messages with group id • Coordinator is unique within a group

16. New Desired Characteristics • Assertion 3: For two nodes i and j • S(i).c = S(j).c if nodes i and j are in the same group and are in one of the states “Normal” or “Reorganization” • S(i).d = S(j).d if both i and j are in “Normal” state and are in the same group

17. Desired Characteristics (contd..) • No requirements for nodes with only one-way communication • Assertion 4: If no failures, election will eventually transform a set of nodes R that have two way communication in any state to a state: • There exists node i with S(i).s =“Normal” and S(i).c = i • Other active nodes j in R have S(j).s = “Normal” and S(j).c = i and S(j).g = S(i).g

18. The Invitation Election Algo. • A node “invites” other nodes to join it in forming a new group • A node may accept or decline an invitation • Make a receiving node form a new group with itself the coordinator and the only member • Objective: to merge groups • Coordinators periodically send “invite” message • Can be used instead of the Bully algorithm

19. Related Work • Scott D. Stoller (2000) • Modifies Bully algorithm to work with crash failures • Points out a flaw and proposes a new specification • Gurdip Singh (1996) • Proposes an algorithm for leader election in the presence of link failures

20. Conclusions • Meaning of an election depends on the possible types of failures • Paper studied elections in two representative failure environments • Postulated assertions that define concept of an election • Presented an election algorithm for each environment