The performance characteristics of a rollback algorithm are analyzed in a simulation experiment. An overview of the operation of the rollback algorithm is presented, followed by a discussion of the simulation model and its parameters. The model, as implemented, consists of data definition, data manipulation command processing, and rollback facilities. The model is parameterized in terms of number of application tasks and amount of data sharing and driven by randomized streams of data manipulation language commands. The simulation experiment applies the model to a CODASYL DBMS in an environment in which the rollback algorithm is initiated by a deadlock detection procedure. The overhead of the rollback algorithm in this environment is projected in terms of CPU utilization and file accesses. The results indicate that the factors having the strongest influence upon the performance of the rollback algorithm are update frequency and the degree of data sharing among programs. The performance projections indicate that the rollback algorithm introduces only a small amount of overhead even in worst case situations.