Check pointing fault tolerance technique record the system state periodically after certain time limit ? or After certain processing limit ?
If it store after certain time limit then what limit is fit for storing the state ?
if it store state after certain processing limit then what limit is fit for the storing the last checkpoint?
Both is possible. The advantage of take a checkpoint in a time-triggered way is to (roughly) know the possible redo time and thus to be better able to tune your overhead. In turn, using a processing limit makes it easier to take a consistent checkpoint.
How frequent take a checkpoint (in terms of time or processing progress) is a design decision that depends on your requirements. At least, you have a tradeoff between the overhead in the fault-free case, and the overhead to restore (redo time). Thus, e.g., if failures happen frequently it makes sense to checkpoint more often than if failures are a more rare event. A rule of thumb is to go for a overhead of less than 10%.
That may change if you have additional requirements. E.g., if you have a deadline you probably have to optimise the probability of meeting the deadline (and provide a correct result). Then it may even worth to change the limit along with the progress (larger chucks at begin, smaller later).
- Badges
- Science method