On the Calculation of the Checkpoint Interval in Run-Time for Parallel Applications

摘要

The growth in the number of components that compose parallel computers increases their fault frequency. Currently, in such systems faults are no longer a rare event but a common problem, thus some sort of fault tolerance should be provided. In general fault tolerance protocols rely on checkpoints. A common question surrounding checkpointing is the definition of the checkpoint interval. Checkpoint interval models define variables which depends on application characteristics, e.g. the time need to take a checkpoint. The use of average values and/or statistical data to define these variables reduces the model's accuracy. In this paper we propose a methodology to define in run-time the variables value needed to calculate the checkpoint interval. While using uncoordinated checkpoint this interval can be defined individually for each process of the parallel application. The variables definition relies on the measuring of the time spent on fault tolerance tasks in run-time. Experimental evaluation shows that the use of our methodology reduces in more than 3% the overhead introduced by fault tolerance while tested applications are running in a faulty environment.