Analysis of a transaction system with checkpointing, failures, and rollback

Abstract

We consider a system where transactions are processed by a single server subject to faults and recovery. A checkpoint is attempted after a fixed number of transactions have been completed, and takes some time to establish. The occurrence of a fault causes a rollback to the last checkpoint, after which all intervening transactions are reprocessed. The system is modelled by a two-dimensional Markov process with one unbounded variable (the number of transactions in the queue), and one bounded variable (the number of transactions processed since the last checkpoint). The joint steady-state distribution of the process, and hence the performance measures of interest, is found by two different methods: generating functions and spectral expansion. The problem of determining the optimal checkpointing parameter is considered. © Springer-Verlag Berlin Heidelberg 2002.