Micro-checkpointing: checkpointing for multithreaded applications

In this paper we introduce an efficient technique for checkpointing multithreaded applications. Our approach makes use of processes constructed around the ARMOR (Adaptive Reconfigurable Mobile Objects of Reliability) paradigm implemented in our Chameleon testbed. ARMOR processes are composed of disjoint elements (objects) with controlled manipulation of element state. These characteristics of ARMORS allow the process state to be collected during runtime in an efficient manner and saved to disk when necessary. We call this approach micro-checkpointing. We demonstrate micro-checkpointing in the Chameleon testbed, an environment for developing reliable distributed applications. Our results show that the overhead ranges from between 39% to 141% with an aggressive checkpointing policy, depending upon the degree to which the process conforms to our ARMOR paradigm