Checkpoint Processing in Distributed Systems Software Using Synchronized Clocks

1
Checkpoint Processing in Distributed Systems
Software Using Synchronized Clocks

• S. NEOGY, A.SINHA and P.K.DAS
• The international conference on Information
  TechnologyCoding and Computing, 2001

2
Introduction

• The method of taking checkpoints in a truly
  distributed manner has been very tricky
• Assume
• The system uses a fault-tolerant hardware
  platform
• A synchronization layer guarantees the
  synchronization of the individual clocks
• There exists a constant Dmax such that at the kth
  resynchronization interval (kgt0) for all correct
  clocks i j, if the logical clocks of the
  processes Pi and Pj be denoted by Cik and Cjk,
  then , Cik(t) - Cjk(t) lt Dmax

3
Checkpoint Model

• Logical clocks of the processes are at most Dmax
  apart from each other
• Each process maintains a message log in the form
  of (message id, sender/receiver id)
• Since communication is synchronous, processes may
  be blocked for send/receive. Checkpointing
  instant may occur during the blocked period.
  Blocked processes take checkpoints after they
  unblock
• An unblocked process takes its checkpoint when
  the checkpointing instant occurs. After taking
  the checkpoint it freezes itself and waits for
  the next clock resynchronization message

4
Algorithm

• Procedure check(Pm)
• if (Pm is not blocked ) then
• take_checkpont(Pm) wait(Pm) until
  clock_synchronization(Pm)
• append the following with the clock
  synchronization message
• i. checkpoint_commit_signal ii.
  message_log
• send clock_synchronization_message to Pk
  where k 0,1,2,..
• resume(Pm)
• else / Pm isblocked /
• wait(Pm) until unblocked
  take_checkpoint(Pm)
• wait(Pm) until clock_synchronization(Pm)
• append the following with the clock
  synchronization message
• i. checkpoint_commit_signal
• ii. message_log
• send clock_synchronization_message to Pk
  where k 0,1,2,..
• resume(Pm)

5
Example(1)
CPik-1
CPik
Xi
Pi
e1ij
e2ij
CPjk
CPjk-1
Pj
Xj
e2ij
e1ij
6
Example(2)
CPik-1
Xi
CPik
Pi
e1ij
e2ij
CPjk
CPjk-1
Pj
Xj
e2ij
e1ij
CPik-1
Xi
CPik
Pi
e1ij
e2ij
CPjk-1
CPjk
Pj
Xj
e2ij
e1ij
7
Conclusion

• Synchronised clock-based checkpointing method
• No central checkpoint coordinator
• Only one checkpoint needs to be stored in the
  stable storage
• The system does not have to roll back more than
  once to restart from a previous consistent state
  in case recovery is required