The MaStA I/O Cost Model and its Validation Strategy

Abstract

Crash recovery in database systems aims to provide an acceptable level of protection from failure at a given engineering cost. A large number of recovery mechanisms are known, and have been compared both analytically and empirically. However, recent trends in computer hardware present different engineering tradeoffs in the design of recovery mechanisms. In particular, the comparative improvement in the speed of processors over disks suggests that disk I/O activity is the dominant expense. Furthermore, the improvement of disk transfer time relative to seek time has made patterns of disk access more significant. The contribution of the MaStA (Massachusetts St Andrews) cost model is that it is structured independently of machine architectures and application workloads. It determines costs in terms of I/O categories, access patterns and application workload parameters. The main features of the model are: Cost is based upon a probabilistic estimation of disk activity, broken down into sequential, asynchronous, clustered synchronous, and unclustered synchronous disk accesses for each recovery scheme. The model may be calibrated by different disk performance characteristics, either simulated, measured by experiment or predicted by analysis. The model may be used over a wide variety of workloads, including those typical of object-oriented and database programming systems. The paper contains a description of the model and illustrates its utility by analysing four recovery mechanisms, delivering performance predictions for these mechanisms when used for some specific workloads and execution platforms. The refinement of I/O cost into the various access patterns is shown to give qualitative predictions differing from those of uniform access time models. Further the results are shown to vary qualitatively between two commercially available configurations. The paper concludes by proposing a validation strategy for the model.KeywordsAccess PatternRecovery MechanismDisk AccessPage TableMaStA ModelThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.