Diffusing Computation

Using Spanning Tree Construction for Solving Leader Election

Root is the leader In the presence of faults,
There may be multiple trees
Multiple leaders

After recovery
There is a unique tree There is a unique leader

Using Spanning Tree Construction for Solving Leader Election

This provides a nonmasking fault-tolerant solution
Eventually, there is one leader During recovery,
The safety specification could be violated because there could be multiple leaders

Designing Masking Fault-Tolerant Solution for Leader Election

Need to ensure that there is at most one leader during recovery Can be achieved if a node makes sure that before it declares itself to be the leader, there is no other leader
If this node were to become the true leader, the tree would be reconstructed to be rooted at that node Check if the tree has been reconstructed
I.e., check if all nodes are in the tree rooted at the current node

How?
When j changes P.j to j
Check if all nodes have set their root value to j? Achieved through diffusing computation

Diffusing Computation
Initiation
The node that initiates the diffusing computation sends the diffusing computation message to all its children

Propagation
A node propagates the diffusing computation when it receives it
To propagate, the node sends the diffusion message to all its children A node can check some predicate (condition) during propagation

Diffusing Computation
Completion
A node completes the diffusing computation iff
All its children complete the diffusing computation All its neighbors have received (propagated) that diffusing computation A node can also check some predicate during completion
The results from this condition are propagated towards the root;

The diffusing computation completes when the root completes the diffusing computation
If the condition checked by the root evaluates to true, we say that the diffusing computation completes successfully Else, we say that the diffusing computation fails/completes unsuccessfully

Using Spanning Tree for Leader Election

When a node is about to become leader
It starts a diffusing Computation If it completes successfully, declare itself to be the leader

Goal of the diffusing computation is to check if the tree is formed at the initiator node
In other words, check if root value of all nodes is equal to that of the initiator

 Issues
Multiple nodes may start diffusing computation
Node with higher ID should win (I.e., its diffusing computation should complete successfully) Node with lower ID should lose.

Failures during diffusing computation

Fail the current diffusing computation

 Upon completing an unsuccessful diffusing computation

Start another one if the node is still likely to be the leader (P.j = j) The diffusing computation may have failed due to faults

Use sequence number to distinguish between different diffusing computations initiated by the same node

 Init
P.j = j
Phase.j = prop Sn.j = newseq() Res.j = true
(Result = result of the diffusing computation) Currently set to true, to be read only after the diffusing computation completes

 Propagation
Root.j = root.(P.j) /\ sn.j sn.(P.j)
If (phase.(P.j) = prop)
Phase.j = prop, res.j = true

Else
Res.j = false // Fail this diffusing computation

 Completion
Phase.j = prop /\ k : k Neighbors.j : root.j = root.k /\ sn.j = sn.k /\ k : k Ch.j : phase.j = comp
res.j = k : k Neighbors.j {j} : res.j Phase.j = j If (P.j = j /\ res.j) Init(j) // Initiate a new diffusing computation

 Aborting Diffusing Computation

Phase.j = comp, res.j = false If (P.j Neighbors.j) res.(P.j) = false

Combining with Tree Algorithm

When j executes tree correction action 1 (red color propagation)
Abort(j)

When j executes tree correction action 2 (changing color to green)

Init(j)

When j changes tree

Abort(j)

Application in Mutual Exclusion

Diffusing Computation
Can be performed in the absence of a tree
Tree is formed during diffusing computation
The initiator sends the first diffusion message If node, say j, receives a diffusion from k then
If this is NOT the first diffusion message Send a reply to k If this is the first diffusion message then Send the diffusion to all neighbors Send a reply to k after replies from ALL neighbors is received A condition can be checked based on the replies received

Diffusing Computation (continued)

The node from which the diffusion message is received for the first time is the parent of that node.
Several problems can be solved by taking appropriate actions during propagation and completion of diffusion
Detecting global state Termination detection

Detecting Global State

I will introduce the problem. A solution is posted on the web, for reference.

Global State
The global state of a system consists of
One local state for each process
Contains all the messages sent and received upto a point in computation

Constraints on how the local state can be collected

Consistency in Global State

Consistent iff
If reception of any message is recorded in the global state then the corresponding send is also recorded

Note that it is possible that the consistent global state records the send of a message but not receive
It implies that the message is in transit

Examples of (in)Consistent Global States

Revisit Dijkstra Safra Termination Detection Algorithm

Note that the token is collecting a global snapshot of the system
Can we see if it is consistent?

2 Phase Termination Detection

Maintain c.j similar to Dijkstra-Safra Termination Detection
But no color variable maintained

Application of Global State Detection

Termination detection Checkpointing and recovery