Roll-forward checkpointing schemes

Abstract

In modular redundant systems, tasks are replicated to achieve fault-tolerance. Checkpointing schemes that exploit replication can achieve better performance than the ones that ignore how the fault detection mechanism is implemented [24]. This Chapter presents two such schemes named Dynamic Roll-Forward Checkpointing Scheme and the Static Roll-Forward Checkpointing Scheme.

In the dynamic scheme for duplex systems, each task is assumed to be executing simultaneously on two processing modules. At each checkpoint, the state of the two modules executing the task is compared for detection of faults. If a fault is detected, instead of the usual roll-back, both the modules continue execution to the next checkpoint interval. The failed checkpoint interval is ‘retried’ on a spare module, which helps in identifying the failed processing module and making its state consistent.

It is demonstrated that this scheme increases the likelihood of a task completing within a specified deadline in spite of transient faults. The dynamic scheme also results in a lower average execution time with a lower variance as compared to the usual duplex roll-back schemes.

The dynamic scheme avoids a roll-back in most cases if the transient faults are independent. However, for correlated faults, it may cause multiple roll-backs. The static scheme is capable of tolerating both independent and correlated faults. In the static scheme for triplex systems, each task is assumed to be executing on three processing modules. At each checkpoint, the state of the three processing modules is compared for detection of faults. Thus, it can tolerate all single faults by masking. In the event of multiple failures, none of the checkpoints match. In that case, various recovery actions are possible depending on the choice of concurrent depth. For initiating a roll-forward action, one of the three processing modules is rolled back to execute the interval that experienced failure, while the other two modules continue execution to the next checkpoint interval. The module that was rolled back helps in identifying the faulty modules and the recovery action continues. In this roll-forward scheme, we do not require any spare modules; thereby avoiding the need for task migration. Simulation results indicate that this scheme outperforms the dynamic scheme in meeting deadlines in the presence of correlated faults. It also results in a lower execution time with lower variance as compared to the static scheme.

Research supported in part by ONR