Effect of update merging on reliable storage performance

Abstract

Performance of a reliable storage subsystem for a centralized database system was studied by simulation. The reliable storage subsystem studied consists of three redundant disk units that are updated one at a time from a consistent database state to another consistent database state. Thus, even if a central processor and/or one disk unit fail simultaneously, at least one disk unit will contain a consistent database state. Redundant multiple disk units allow simultaneous processing of multiple read operations. Our simulation result shows that when all transactions are read-only transactions, the throughput of a three-unit system is 2.5 times higher than that of a single-unit system. On the other hand, redundant disk units slows down write operation because each disk unit must be updated for each virtual page updated. The major result in this paper shows that this adverse effect can be mitigated if updates of multiple transactions are merged and if an efficient disk scheduling policy (e.g., CSCAN) is employed. When the ratio of write operations to read operations is between 10 to 30%, the throughput of the three-unit system is still 89 to 144% higher than that of the single-unit system, which is not reliable. This result is significant when we consider that the performance of a reliable storage subsystem based on logging cannot exceed that of a single-unit storage subsystem.