1. Presented by Rukmini and Diksha Chauhan Virginia Tech 2 nd May, 2007 Movement-Based Checkpointing and Logging for Recovery in Mobile Computing Systems Sapna E.George, Ing-Ray Chen & Ying Jin

2. Agenda Related work Mobile Computing System Proposed Movement-based Checkpointing and Logging Recovery Schemes Performance Analysis Conclusion

3. Properties of Mobile Computing Inherent properties Host Mobility Disconnections Wireless bandwidth Limitation Battery Life Storage Hardware failure Software Failure Motivation Propose an efficient failure recovery scheme

4. Distributed Systems Fault-tolerance schemes Logging Checkpointing Rollback Recovery Definition Domino Effect inter-process dependencies - cascading rollbacks Asynchronous Recovery

5. Related Work Acharya et al. in [1] describes a distributed uncoordinated checkpointing scheme, where multiple MHs can arrive at a global consistent checkpoint without coordination messages. The paper does not describe how failure recovery is achieved nor does it address the issue of recovery information management in the face of MH movement.

6. Underlying Model

7. Basic Definitions Mobile Mobile Host(MH) Mobile support Systems(MSS) Infrastructue machines High speed Static wired n/w

8. Basic Definitions Cell Local MSS Communication Between MH and MSS-Constraints Process of Communication between MH’s Two one-hop wireless transmissions Arbitrary hops

9. Basic Definitions Handoff Instantaneously Process MH crosses a cell boundary MH disconnect(MSS1) voluntarily from network to conserve power and reconnect(MSS2) at a later time. MH sends the ID of MSS1 to the new MSS2-initiates handoff procedures.

10. Processes and States Three States Normal Execution Application-related Computation Sending or receiving messages Logging Save Recovery Write Event Message received from other MH or server User Input or Local Computation

11. Movement-Based Checkpointing & Message Logging Checkpoints after a certain number of host migrations across cells rather than periodically. Recovery Scheme Combines independent checkpointing and optimistic message logging enabling asynchronous recovery of a MH upon failure. Application recovery mechanisms - optimize recovery cost (failure-free operational cost), recovery time Storage requirements for recovery related information

12. Movement-Based Checkpointing & Message Logging Scheme uses distance or number of handoffs Parameter to trigger information consolidation MH crosses a distance threshold from the location of the latest checkpoint, the recovery information is collected and transferred to the MH’s local MSS. Recovery protocol – proactively controls no. of checkpoints and logs by movement-based checkpointing strategy additional overhead of unnecessary checkpoints and log consolidation during failure-free operation is avoided.

13. Checkpointing & Message Logging m is a f (user’s mobility rate, the failure rate and log arrival rate ) – Adaptation to user and Application behaviour

14. Movement-based Checkpointing and Logging MH –Stored variables cp_seq -stores the sequence number of the latest checkpoint and cp_loc -stores the ID of the MSS that has recorded the latest checkpoint. MSScp- Latest MSS Handoff_counter to 0 MSSlogs (log_set) - IDs of MSSs Activity Checkpoints Logging

15. Independent Recovery Independent – without Coordination with other hosts. Recovery process MH sends MSS cp_seq, cp_loc and log_set MSS initiates (requests) data collection MSS compiles Logs into list ordered by time Checkpoints Once recovery is completed successfully, a checkpoint of the current state is taken and sent to the MSS and the local variables are reset.

16. Storage Management at MSS MH’s Disk Unstable Limited MSS’s Disk Stable storage Considerably large storage at MSSs –depleted 1. Temporarily halt –Perform Garbage Collection 2. Alternative Storage 3. Deleting outdated recovery Information

17. SPN Model Parameters

18. SPN Model

19. Transition firing rate Checkpoint Rate of MH During checkpointing: (a)MH takes a snapshot of its current state (b)MH sends the checkpoint to the current MSS through the wireless channel. (c)The MSS stores it in its stable storage. where is the time required to transmit a checkpoint through wireless link.

20. Transition Firing Rate Recovery Rate of MH i.e, inverse of Recovery Time Recovery Time includes : (a)time to send recovery information requests to the MSSs storing the latest checkpoint and all logs since the latest checkpoint (b) time to transmit the latest checkpoint from the MSS where it is stored (MSScp) to the MSS in which the MH has recovered (MSSrec) through the wired network and through the wireless channel to the MH

21. Transition Firing Rate c) time to transmit all the logs from the respective MSSs where they are located (MSSlogs) to the MSSrec through the wired network and through the wireless channel to the MH and (d) time to rollback to the last checkpoint and apply all the logs at the MH.

22. Variables & Represents the number of MSSs storing logs.At most its value is the number of handoffs before failure, i.e. i Represents average hop count between MSScp and MSSrec.

23. Recovery Time Time Spent on Recovery Requests: Time spent on transmitting the latest checkpoint to MH:

24. Recovery Time contd.. Time spent to transmit the logs to MH: where n is the number of log entries since the last checkpoint Time spent to rollback to the last checkpoint and apply the logs: Total Recovery time after i movements:

25. Recovery Cost per failure The SPN model’s underlying Markov model has 2M+1 states. The average recovery time per failure is given by: The total failure-free operations cost(or time spent on checkpointing and logging before failure) is given by: where denotes the number of checkpoints before failure and denotes the number of log entries before failure.

26. Recovery Cost per failure contd… Total Cost of Recovery per failure is the weighted sum of the average recovery time and the total time spent on the checkpointing and logging per failure and is given by: where w1 and w2 are the weights associated with recovery time and failure-free operation cost. This paper uses w1 = w2 = 0.5 to account for the situation where is equally proportional to and

27. Recovery Probability The recovery probability is defined as the probability that recovery time is less than or equal to T

28. Results and Analysis The SPN model was implemented and analyzed using the SPNP s/w The following parameter values were kept constant in all the runs. size of a log entry is 50B, size of a checkpoint is 2000B, bandwidth of the wired network is 2Mbps, ratio of bandwidth of wireless to wired network (r) is 0.1, Telog is 0.0001s. Tlog_w is 0.002s and Tckp_w is 0.08s. Model parameters such as mobility rate, log arrival rate, failure rate, and movement threshold were varied across runs

29. Results and Analysis contd… Recovery Probability vs. Recovery Time.

30. Results and Analysis contd… Recovery Probability vs. Log Arrival Rate.

31. Results and Analysis contd… Recovery Probability vs. Failure Rate.

32. Results and Analysis contd… Recovery Probability vs. Movement Threshold.

33. Results and Analysis contd… Recovery Time vs. Movement Threshold.

34. Results and Analysis contd… Determining Optimal Movement Threshold that minimizes Recovery Cost Per Failure.

35. Applicability Results can be applied in the following manner: Build a Table at static time covering possible parameter values of the mobility rate and failure rate of the MH and log arrival rate of the mobile applications List the optimal M value to minimize the recovery cost per failure for each parameter set. Select optimal M dynamically at runtime based on the measured rates to minimize the recovery cost per failure.

36. Summary Implemented movement-based checkpointing and logging scheme which checkpoints after M movements (mobility handoffs) as compared to current approaches where checkpoints are taken periodically. A performance model developed based on stochastic Petri nets to identify the optimal M, given the failure, mobility and log arrival rates, to minimize the cost of recovery per failure. The results of performance analysis and the sensitivity of recoverability to the various parameters were shown

37. Future Work To analyze and compare the proposed approach to existing approaches, in terms of the gain achieved over the use of constant periodic checkpointing. To extend the proposed work to MIPv6 environments.

38. QUESTIONS ??