Partial Order Based Runtime Recovery Intended For Highly Available Distributed Applications
Ching Wei Su · Spectrum Research Repository (Concordia University) · 2011
This thesis develops a checkpoint based runtime rollback recovery techniqueintended to be used with highly available distributed applications whose correct expected behavior is specified by the application developer through a partially ordered multiset (POMSET) of application events.Checkpoint based rollback recovery techniques for distributed applications typically store the state of all the application processes and application events (called the global checkpoints) in persistent store at periodic time intervals.When a runtime failure is detected, the application is rolled back to an appropriate correct past state in its execution using the saves checkpoints.Such techniques do not know the correct application behavior and hence have to store large amount of state requiring significant amount of persistent storage and recovery time.The idea behind this thesis is that knowing the correct expected behavior of the application, only the checkpoints necessary to ensure runtime recovery of the application can be identified and stored thereby making the recovery much more efficient.The application developer specifies the correct expected behavior of the distributed application through a POMSET of application events which is stored as a tree.The developed runtime recovery technique identifies the nodes of the POMSET tree at