Proposta de métricas para o garbage collector assíncrono RPGC

Abstract

Distributed simulations require time synchronization strategies to prevent simulation events from being processed out of timestamp order. Conservative strategies prevent time violation. Optimistic strategies allow time violations to happen but rollback the simulation to a previously safe state, denominated checkpoint. As the simulation advances, the amount of checkpoints tend to grow, which results in memory usage increase. Optimistic synchronization strategies are classified in coordinated, uncoordinated e communication induced. In coordinated synchro- nization, removing checkpoints is a trivial task. In uncoordinated and communication-induced synchronization strategies there are garbage collection algorithms that are executed periodically in the removal of checkpoints. RPGC is an asynchronous garbage collector that makes use of metrics to determine the amount of checkpoints that should remain in the simulation. However, the existing metrics used tend to keep checkpoints that are not going to be used in future roll- back operations. We present the implementation of one new metric and one new parameter – rollback frequency and the growth percentage of the largest rollback, respectively. Satisfactoy results were obtained in improving the precision of the estimates. Furthermore, some fragilities were identified in the original work. The implementation was integrated into the Distributed Co-Simulation Backbone, a distributed simulation architecture