Lecture 17: Voting Protocol

These topics are from Chapter 13 (Fault Tolerance) in Advanced Concepts in OS

Topics for Today

• voting algorithms

Voting Protocols

• replicated data, at multiple sites
• each site has some number of votes
• access to replicated data requires a majority of votes
• votes determine which version is the current one

Static Voting

• replicated data, at multiple sites
• each file access requires obtaining a lock
• reader-writer locks are supported
• every site has a lock manager
• every file has a version number = number of changes made
• each replica has some number of votes
• vote allocation is on stable storage
• reads and write require a quorum

Static Voting

For a read or write request initiated by site i:

• issue Lock_Request to local lock manager
• local lock manager eventually grants request and then sends Vote_Request to all sites

at site j:

• on receipt of Vote_Request from i, issue Lock_Request to local lock manager
• if the lock request is granted by the local lock manager, send the version number VN_j of its replica and the number of votes V_j of its replica to site i

at site i:

• after votes are in, perform quorum test

Read Quorum Test

Where P is the set of sites that replied.

Write Quorum Test

Where M = max{VN_j | j Î P} is the largest version number reported in the vote, and
Q = {j Î P | VN_j = M } includes only the votes that correspond to that version number

Voting Algorithm (continued)

at site i:

• If quorum test fails, issue Release_Lock to local manager and all sites in P that returned positive votes
• If quorum test succeeds, check whether local copy is current. If not, obtain a fresh copy from another site.
• For a read, just use the local copy.
• For a write, update the local copy.
  Then update VN_i and send the updates and VN_i to all the sites in Q.
• Issue a Release_Lock to local manager and all the sites in P

at other sites:

• on receiving update messages, update own local copies
• on Release_Lock, release all local locks

Vote Assignment

If v is the total number of votes, we want to choose r and w such that
r + w > v and
w > v/2

Why?

Consequences

• no obsolete copies are updated due to a write operation
• the current local replicas have at least w votes
• every read quorum and write quorum overlap by at least one site
• there cannot be simultaneous writes on distinct sets of replicas

How many temporary site failures can we tolerate?

What happens when a site comes back on line after failing?

What happens if the network is partitioned?

Tuning Example

If r=1 and w=5, the read access time is 75ms and the write access time is 750 ms. Any single site failure will prevent writes.

If r=3 and w=3, the access times are unchanged, but writes are still possible with a single site failure.

If site 4 is more reliable, we can further improve reliability by readjusting the votes as follows.

Dynamic Voting Protocols

• Change the set of sites that can form a majority
• Change the distribution of votes

Dynamic Vote Reassignment

• number of votes per site changes
• two kinds:
• group consensus on new assignment
• autonomous reassignment, ratified by majority of sites

What are the strengths and weakenesses of each?

Autonomous Vote Reassignment

• each site i has vector V_i representing its belief of the global vote assignment
• V_i[j] is how many votes i thinks j is entitled to have
• each site i has version-number vector N_i
• N_i[j] is the version number of V_i[j]
• each site i has vector v_i representing the votes it has seen
• v_i[j] is how many votes i sees j is trying to cast

Vote Increasing Protocol

When site i wants to increase V_i[i]:

• send V_i and N_i along with new vote value x to all communicating sites
• wait for a majority of sites to respond
• if a majority is collected, update V_i[i] to the new value and increment N_i[i].

When site j receives a vote-increasing request from site i with V_i, N_i, and x:

• V_j[i] = x
• N_j[i] = N_i[i] + 1

Vote Decreasing Protocol

When site i wants to decrease V_i[i]:

• set V_i[i] to the new value
• increment N_i[i]
• send V_i and N_i to the other sites

When site j receives a vote-decreasing request from site i with V_i and N_i:

• V_j[i] = V_i[i]
• N_j[i] = N_i[i]

Vote Collecting Protocol

• for each reply V_j and N_j received by site i:
  • v_i[j] = V_j[j]
  • if V_j[j] > V_i[j] or (V_j[j] < V_i[j] and N_j[j] > N_i[j]) then
       V_i[j] = V_j[j]; N_i[j] = N_j[j]
    end if;
• if site j did not respond to site i:
  • Find k Î G such that N_k[j] = max {N_p[j] | p Î G}, where G is the set of all sites that replied to i. That is, find the site that has the latest information on the votes assigned to site j.
  • v_i[j] = V_k[j]; V_i[j] = V_k[j]; N_i[j] = N_k[j]

Deciding the Outcome

Let K be the set of all sites, and G be the set of sites that responded to the ballot.

Site i has a majority iff RCVD > TOT/2.

Vote Increasing Policies

The above all leaves open the question of when a site should try to increase or decrease its vote.

This is normally done in response to detection of an apparent failure.

• overthrow technique -- one site increases its vote
• alliance technique -- all active sites increase their votes