Snapshots

1
Snapshots

• Goal Capture the current configuration of a
  distributed system
• Motivation
• analyze the properties of the computation
• for stable properties, i.e termination,
  deadlock, loss of tokens, non-reachability of
  objects (garbage collection)
• failure recovery
• debugging
• Problems
• WHAT is a snapshot?
• HOW to construct it?

2
Preliminaries

• Local snapshot
• snapshot state (local state, communication
  history)
• preshot and postshot events
• sentp,q , recvp,q
• Global snapshot
• collection of local snapshots
• Feasible (Global) snapshot
• for all neighbours p,q recvp,q is a subset
  of sentp,q

3
Preliminaries II

• Consistent Cut L
• if an event is in L then also all events that
  causally precede it are in L
• Meaningful snapshot (in computation C)
• a snapshot that captures a configuration that
  occurs in some execution of the computation C
• Lemma Let S be a snapshot and let L be the cut
  implied by S. The following three statements are
  equivalent
• S is feasible
• L is a consistent cut
• S is meaningful

4
Snapshot Algorithms

• It is sufficient to ensure that
• local snapshot is taken in each process
• no postshot message is received in a preshot
  event
• Chandy Lamport Algorithm
• assumes FIFO channels
• Lai Yang Algorithm
• does not assume FIFO channels, but piggybacking
  must be possible
• TAYLOR If neither FIFO, nor piggybacking is
  available, each snapshot algorithm must be
  inhibitory.

5
Chandy Lamport Algorithm
Idea flood the network, take snapshot when
message received Spontanneous wake-up take
local snapshot send(take_snapshot) to all
neighbours Upon receiving take_snapshot
message if no snapshot has been taken yet
then take local snapshot send(take_snapshot)
to all neighbours except the one from which
the message arrived Complexity (overhead) 2E
take_snapshot messages Correctness follows from
FIFO, as no postshot message can overtake the
take_snapshot message
6
Lai Yang Algorithm

• Idea each node p maintains variable takenp,
  whether it has taken the snapshot or not. This is
  appended to each message sent.
• A node takes snapshot either spontaneously or
  when it receives the first message with taken
  true. (The snapshot is taken before processing
  the contents of that message.)
• Correctness
• no postshot message is processed before a local
  snapshot is taken
• no assurance that each node will take local
  snapshot
• initial flooding of control messages may be
  necessary
• Complexity if no initial flooding need, then
  just one biggybacked bit on each message,
  otherwise the cost of initial flooding

7
Using Snapshots Deadlock Detection

• Model of the basic algorithm
• when a process p becomes blocked, it sends
  requests to nodes Reqp
• predicate Freep defines the subsets of Reqp from
  which the grant messages are sufficient to free p
• a node collects requests in a set Pendp, it can
  grant request (send back grant message) only if
  it is active
• request and grant numbers and attached to the
  messages, so that a node can discard obsolete
  grant messages
• Deadlock
• a node p is alive if there is a reachable
  configuration in which p is active
• p is deadlocked if it is not alive
• a configuration is deadlocked if there is a
  deadlocked process

8
Deadlock Detection

• Requirements
• non-interference dont influence the basic
  algorithm
• liveness deadlock will be detected, if present
• safety algorithm detects deadlock only if there
  indeed is deadlock
• usually we also to report which nodes are
  deadlocked

9
Global Marking Algorithm

• Overall structure
• repeatedly take a snapshot and check it for
  deadlock
• Checking idea
• simulate propagations of grant from the snapshot
  on, assuming every active process grants access
  to all requests it received
• Algorithm
• At an active node
• Send alive message to all node which sent
  requests to me.
• At a blocked node p
• Receive alive messages. If the predicate Freep
  is satisfied, become active.