Fault Tolerance for Real-Time Systems: Analysis and Optimization of Roll-back Recovery with Checkpointing Nikolov, Dimitar

Increasing soft error rates in recent semiconductor technologies enforce the usage of fault tolerance. While fault tolerance enables correct operation in the presence of soft errors, it usually introduces a time overhead. The time overhead is particularly important for a group of computer systems referred to as real-time systems (RTSs) where correct operation is defined as producing the correct result of a computation while satisfying given time constraints (deadlines). Depending on the consequences when the deadlines are violated, RTSs are classified into soft and hard RTSs. While violating deadlines in soft RTSs usually results in some performance degradation, violating deadlines in hard RTSs results in catastrophic consequences. To determine if deadlines are met, RTSs are analyzed with respect to average execution time (AET) and worst case execution time (WCET), where AET is used for soft RTSs, and WCET is used for hard RTSs. When fault tolerance is employed in both soft and hard RTSs, the time overhead caused due to usage of fault tolerance may be the reason that deadlines in RTSs are violated. Therefore, there is a need to optimize the usage of fault tolerance in RTSs. To enable correct operation of RTSs in the presence of soft errors, in this thesis we consider a fault tolerance technique, Roll-back Recovery with Checkpointing (RRC), that efficiently copes with soft errors. The major drawback of RRC is that it introduces a time overhead which depends on the number of checkpoints that are used in RRC. Depending on how the checkpoints are distributed throughout the execution of the job, we consider the two checkpointing schemes: equidistant checkpointing, where the checkpoints are evenly distributed, and non-equidistant checkpointing, where the checkpoints are not evenly distributed. The goal of this thesis is to provide an optimization framework for RRC when used in RTSs while considering different optimization objectives which are important for RTSs.