TRADEOFFS BETWEEN CENTRALIZED AND GEOGRAPHICALLY DISTRIBUTED TRANSACTION PROCESSING SYSTEMS

In a geographically distributed transaction processing system, the database is partitioned among the regional processing systems, and the shipping of database requests across a long haul network may be required to access remote databases. This is in contrast to a centralized (or locally distributed) system in which the transaction input and output is shipped from the user workstations to a centrally located complex. These two approaches are compared, and it is found that the geographically distributed database approach offers a marginal improvement in transaction response time if the number of remote calls per transaction is quite small, but will have a much poorer response time than the centralized system otherwise. Similarly, the geographically distributed approach allows a small saving in communications cost if the number of remote calls per transaction is small, but otherwise will cost significantly more. A performance model is developed to quantify the trade-offs, to determine the cross-over point between the two approaches, and to show the sensitivity to various parameters. It is demonstrated that the centralized (or locally distributed) database approach is more robust to variations in the workload and transaction character-istics. Finally, it is observed that to create a viable distributed system from an existing centralized system would require a restructuring of the entire application.