Improvements to a Roll-Back Mechanism for Asynchronous Checkpointing and Recovery

Gupta, Rahimi and Yang recently proposed a novel recovery algorithm for distributed systems in which checkpoints are taken asynchronously [1]. A checkpoint taken by a process is a snapshot of its local state, stored in a stable storage, so that the process can roll back to it, if this becomes necessary. The start of a process is also one of its checkpoints. Asynchronous checkpointing means that processes take their checkpoints independently. A failure in a distributed system in principle requires that all its constituent processes roll back, to a global state from which the system can resume its operation as if it had started from it, i.e., to a globally consistent set of local checkpoints (GCSLC), ideally to the so called maximum GCSLC, in which every local checkpoint is as recent as possible. In the case of asynchronous checkpointing, when a failure occurs, the processes have yet to find the maximum GCSLC. They do that by running a checkpoint coordination algorithm (CCA). Alternatively, the processes might agree to restart from the GCSLC which they currently treat as the starting state of the system, i.e., from the current recovery line. This is the most recently computed maximum GCSLC, initially the actual starting state of the system. [1] suggests that the processes occasionally initiate a CCA just for advancing the recovery line. In this paper, we demonstrate that, because of a subtle logical flaw, the CCA of [1] sometimes returns a checkpoint set which is not globally consistent. We correct the flaw and also suggest some other improvements. The rest of the paper is organized as follows. In the next section, we describe the system and the testing of checkpoint consistency and give a brief outline of the CCA of [1]. In Section 3, we explain and correct the flaw. In Section 4, we suggest several other improvements of the CCA. A detailed specification of the improved CCA is given in Section 5. Section 6 suggests that checkpointing and recoveryline advancing in the absence of failures should be more flexible. As the proposed CCA correction increases the usage of the local stable storages, we in Section 7 suggest how to organize them. Section 8 comprises a discussion and conclusions.