Software schemes of reconfiguration and recovery in distributed memory multicomputers using the actor model

Abstract

Ideally, a multicomputer system should cope with a processor failure by reconstructing itself-and the application running on itself-in order to maintain the available computational power of the remaining processors. We discuss the continuance of running applications through permanent processor failures. We take advantage of the characteristics of the actor model of parallel computation and dynamically checkpoint the activity of the application. Consequently, the runtime system is able to continue an application through multiple nonconcurrent processor failures. We have implemented our techniques through modifications of the runtime system of the parallel language Charm on an Intel iPSC/s hypercube. After discussing the theory and implementation, we give measurements of overhead due to fault tolerance for a number of applications and demonstrate continuance of the applications after injection of one or more faults.<>