1. Chapter 2 Wenbing Zhao Department of Electrical and Computer Engineering Cleveland State University wenbing@ieee.org Building Dependable Distributed Systems Building Dependable Distributed Systems, Copyright Wenbing Zhao 1

2. Wenbing Zhao Outline Checkpointing and logging  System models  Checkpoint-based protocols Uncoordinted checkpointing Coordinated checkpointing  Logging-based protocols Pessimistic logging Optimistic logging Causal logging

3. Building Dependable Distributed Systems, Copyright Wenbing Zhao Wenbing Zhao Checkpointing and Logging: Checkpointing and logging are the most essential techniques to achieve dependability  By themselves, they provide rollback recovery  They are used for more sophisticated dependability schemes Checkpoint: a copy of the system state  Can be used to recover the system to the state when the checkpoint was taken Checkpointing: the action of taking a copy of the system state, typically periodically Logging: log incoming/outgoing messages, etc.

4. Building Dependable Distributed Systems, Copyright Wenbing Zhao Wenbing Zhao Rollback Recovery vs. Rollforward Recovery

5. System Models Distributed system model Global state: consistent, inconsistent Distributed system model redefined Piecewise deterministic assumption Output commit Stable storage Building Dependable Distributed Systems, Copyright Wenbing Zhao 5

6. System Models Distributed system  A DS consists of N processes  A process may interact with other processes only by means of sending and receiving messages  A process may interact with another process within the DS, or a process in the outside world Fault Model: fail stop Building Dependable Distributed Systems, Copyright Wenbing Zhao 6

7. System Models Process state  Defined by its entire address space in OS  Relevant info can be captured by user-supplied APIs Global state  The state of the entire distributed systems  Not a simple aggregation of the states of the processes Building Dependable Distributed Systems, Copyright Wenbing Zhao 7

8. Capturing Global State Global state can be captured using a set of individual checkpoints Inconsistent state: checkpoints reflects message received but not sent Building Dependable Distributed Systems, Copyright Wenbing Zhao 8

9. Capturing Global State: Example P0: bank account A, P1: bank account B m0: deposit $100 to B (after A has debited A) P0 takes checkpoint C0 before debit op P1 takes checkpoint C1 after depositing $100 Scenario: P0 crashes after sending m0, and P1 crashes after taking C1 If the global state is reconstructed based on C0 and C1, it would appear that P1 got $100 from nowhere Typos: p17, section 2.1.2 & p18, example 2.1: Figure 2.2(a) => Figure 2.2 (c) Building Dependable Distributed Systems, Copyright Wenbing Zhao 9

10. Capturing Global State: Example P0 takes checkpoint C0 after sending m0 (reflect debit of $100) P1 takes checkpoint C1 after depositing $100 Dependency of P0 and P1 is captured by C0 and C1 Global state can be reconstructed based on C0 and C1 correctly Typos: p19, example 2.1: Figure 2.2(b) => Figure 2.2 (a) Figure 2.2(c) => Figure 2.2 (b) Building Dependable Distributed Systems, Copyright Wenbing Zhao 10

11. Capturing Global State: Example P0 takes checkpoint C0 after sending m0 (reflect debit of $100) P1 takes checkpoint C1 before receiving m0 but after sending m1 P2 takes checkpoint C3 before receiving m1 If using C0, C1, C3 to reconstruct global state, it would appear that m0 is sent but not received  Debit $100 from A, but not deposited to B However, the reconstructed global state is still regarded as consistent because this state could have happened:  m0 and m1 are still in transit => channel state Building Dependable Distributed Systems, Copyright Wenbing Zhao 11

12. Distributed System Model Redefined A distributed system consists of the following:  A set of N processes Each process consists of a set of states and a set of events One of the states is the initial state The change of states is caused by an event  A set of channels Each channel is a uni-directional reliable communication channel between two processes The state of a channel consists of the set of messages in transit in the channel A pair of neighboring processes are connected by a pair of channels, one in each direction.  An event (such as the sending or receiving of a message) at a process may change the state of the process and the state of the channel it is associated with, if any Building Dependable Distributed Systems, Copyright Wenbing Zhao 12

13. Back on the Global State Example Global state consists of  C0, C1, and C2  Channel state from P0 to P1: m0  Channel state from P1 to P2: m1 Building Dependable Distributed Systems, Copyright Wenbing Zhao 13

14. Piecewise Deterministic Assumption Using checkpoints to restore system state (after a crash) would mean that any execution after a checkpoint is lost Logging of events in between two checkpoints would ensure full recovery Piecewise deterministic assumption:  All nondeterministic events can be identified  Sufficient information (referred to as determinant) that can be used to recreate the event deterministic must be logged for each event  Examples: receiving of a message, system calls, timeouts, etc.  Note that the sending of a message is not a nondeterministic event (it is determined by another nondeterministic event or the initial state) Building Dependable Distributed Systems, Copyright Wenbing Zhao 14

15. Output Commit Once a message is sent to the outside world, the state of the distributed system may be exposed to the outside world Should a failure occur, the outside world cannot be relied upon for recovery Output commit problem: To ensure that the recovered state is consistent with the external view, sufficient recovery information must be logged prior to the sending of a message to the outside world. Building Dependable Distributed Systems, Copyright Wenbing Zhao 15

16. Stable Storage Checkpoints and events must be logged to stable storage that can survive failures for recovery Various forms of stable storage  Redundant disks: RAID-1, RAID-5  Replicated file systems: GFS Building Dependable Distributed Systems, Copyright Wenbing Zhao 16

17. Checkpoint-Based Protocols Uncoordinated protocols Coordinated protocols Building Dependable Distributed Systems, Copyright Wenbing Zhao 17

18. Uncoordinated Checkpointing Uncoordinated checkpointing: full autonomy, appears to be simple. However, we do not recommend it for two reasons  Checkpoints taken might not be useful to reconstruct a consistent global state Cascading rollback to the initial state (domino effect)  To enable the selection of a set of consistent checkpoints during a recovery, the dependency of checkpoints has to be determined and recorded together with each checkpoint Extra overhead and complexity => not simple after all

19. Cascading Rollback Problem Last checkpoint: C 1,1 by P1, before P1 crashed Cannot use C 0,1 at P0 because it is inconsistent with C 1,1 => P0 rollbacks to C 0,0 Cannot use C 2,1 at P2 because it fails to reflect the sending of m6 => P2 rollbacks to C 2,0 Cannot use C 3,1 and C 3,0 as a result => P3 rollbacks to initial state

20. Cascading Rollback Problem The rollback of P3 to initial state would invalidate C 2,0 => P2 rollbacks to initial state P1 rollbacks to C 1,0 due to the rollback of P2 to initial state This would invalidate the use of C 0,0 at P0 => P0 rollbacks to initial state The rollback of P0 to initial state would invalidate the use of C 1,0 at P1 => P1 rollbacks to initial state

21. Tamir and Sequin Global Checkpointing Protocol One of the processes is designated as the coordinator Others are participants The coordinator uses a two-phase commit protocol for consistency on the checkpoints  Global checkpointing is carried out atomically: all or nothing  First phase: create a quiescent point of the distributed system  Second phase: ensure the atomic switchover from old checkpoint to the new one

22. Tamir and Sequin Global Checkpointing Protocol Control messages for coordination  CHECKPOINT message: initiate a global checkpoint & to create quiescent point  SAVED message: to inform the coordinator that local checkpoint is done by participant  FAULT message: a timeout occurred, global checkpointing should abort  RESUME message: to inform participants that it is time to resume normal operation CHECKPOINT certificate: keep track if received it from each incoming channel  Certificate complete: when a CHECKPOINT msg is received from every incoming channel

23. Tamir and Sequin Global Checkpointing Protocol Sending Control messages  CHECKPOINT message: send to every outgoing channel  SAVED message: only to upstream link, i.e., the process from which one receives the CHECKPOINT msg the first time  FAULT message: If originated from the process, send to every outgoing channel If received from a process, send to all outgoing channel except the one that connects to the process from which it receives the FAULT msg  RESUME message: For the coordinator (msg originator): send to all outgoing channels For a participant, send to all outoing channel except the one that connects to the process from which it receives the RESUME msg.

24. Tamir and Sequin Global Checkpointing Protocol Typos: p24, figure 2.4, p25, figure 2.5 Final state machine => Finite state machine

25. Tamir and Sequin Global Checkpointing Protocol SAVED: send to up stream node

26. Tamir and Sequin Global Checkpointing Protocol: Example P0 channel state: m0 P1 channel state: m1 P2 channel state: empty

27. Tamir and Sequin Global Checkpointing Protocol: Proof of Correctness The protocol produces consistent global state Proof: a consistent global state consists of only two scenarios:  All msgs sent by one process prior to its taking a local checkpoint have been received prior to the other process taking its local checkpointing This is the case if no process sends any msg after the global checkpoint is initiated  Some msgs sent by one process prior to its taking a local checkpoint might arrive after the other process has checkpointed its state, but they are logged for replay Msgs received after the initiation of global checkpointing are logged, but not executed, ensuring this property Note that if a process fails, the global checkpointing would abort

28. Chandy and Lamport Distributed Snapshot Protocol CL snapshot protocol is a nonblocking protocol  TS checkpointing protocol is blocking  CL protocol is more desirable for applications that do not wish to suspect normal operation  However, CL protocol is only concerned how to obtain a consistent global checkpoint  CL Protocol: no coordinator, any node may initiate a global checkpointing Data structure  Marker message: equivalent to the CHECKPOINT message  Marker certificate: keep track to see a marker is received from every incoming channel

29. CL Distributed Snapshot Protocol

30. Example P0 channel state: m0 (p1 to p0 channel) P1 channel state: m1 (p2 to p1 channel) P2 channel state: empty

31. Comparison of TS & CL Protocols Similarity  Both rely on control msgs to coordinate checkpointing  Both capture channel state in virtually the same way Start logging channel state upon receiving the 1 st checkpoint msg from another channel Stop logging channel state after received checkpoint on the incoming channel  Communication overhead similar

32. Comparison of TS & CL Protocols Differences: strategies in producing a global checkpoint  TS protocol suspends normal operation upon 1 st checkpoint msg while CL does not  TS protocol captures channel state prior to taking a checkpoint, while CL captures channel state after taking a checkpoint  TS protocol more complete and robust than CL Has fault handling mechanism

33. Log Based Protocols Work might be lost upon recovery using checkpoint- based protocols By logging messages, we may be able to recover the system to where it was prior to the failure System mode: the execution of a process is modeled as a set of consecutive state intervals  Each interval is initiated by a nondeterministic state or initial state  We assume the only type of nondeterministic event is receiving of a message

34. Log Based Protocols In practice, logging is always used together wit checkpointing  Limits the recovery time: start with the latest checkpoint instead of from the initial state  Limits the size of the log: after taking a checkpoint, previously logged events can be purged Logging protocol types:  Pessimistic logging: msgs are logged prior to execution  Optimistic logging: msgs are logged asynchronously  Causal logging: nondeterministic events that not yet logged (to stable storage) are piggybacked with each msg sent For optimistic and causal logging, dependency of processes has to be tracked => more complexity, longer recovery time

35. Pessimistic Logging Synchronously log every incoming message to stable storage prior to execution Each process periodically checkpoints its state: no need for coordination Recovery: a process restores its state using the last checkpoint and replay all logged incoming msgss

36. Pessimistic Logging: Example Pessimistic logging can cope with concurrent failures and the recovery of two or more processes

37. Benefits of Pessimistic Logging Processes do not need to track their dependencies  Logging mechanism is easy to implement and less error prone Output commit is automatically ensured No need to carry out coordinated global checkpointing  By replaying the logged msgs, a process can always bring itself to be consistent with other processes Recovery can be done completely locally  Only impact to other processes: duplicate msgs (can be discarded)