Abstract:

Abstract High performance computing have a high number of constituent components used to facilitate data movement. Key characteristics of these systems include parallel processing, large memory, multiprocessor or multimode communication, and parallel file systems. Though they can turnaround computing in scenarios that need maximum processing power, HPCs face many challenges, key among them being fault tolerance. Today, most applications deal with faults by noting checkpoints frequently. Whenever a fault occurs, all the processes are terminated, and the task is loaded once again from the last checkpoint. Most applications deal with faults by noting checkpoints frequently. Whenever a fault occurs, all the processes are terminated, and the task is loaded once again from the last checkpoint. Key fault tolerance techniques used on HPC applications (reactive and proactive) were evaluated in this paper. Reactive protocols discussed include checkpointing/ restarting, replication, retry, and SGuard, while proactive techniques include preemptive migration, software rejuvenation, and self-healing strategy. As seen from the discussion on the drawbacks of each approach, efficient management of faults can best be achieved by using a hybrid system applying proactive and reactive measures simultaneously.

Keywords: