Fault tolerant dynamic agent systems

Abstract

Partial system snapshots reduce the cost per node to only depend on the size of the connected group instead of the size of the full system. These groups can be determined during system operation by using the communication patterns between nodes. The number of nodes that must rollback after a failure is limited to the size of these snapshot groups, reducing the work lost. These changes to snapshot algorithms are necessary because the cost per node for a snapshot increases and the expected time between failures decreases as the size of the system grows.