High Throughput Distributed Database Systems

本研究的目的是开发分布式数据库系统中事务的并发控制算法，以实现高吞吐量和低响应时间。数据复制是为此目的而提出的一种机制。不幸的是，事务行为可串行化的要求强制在副本站点之间进行相当大的同步，从而对吞吐量和响应时间产生不利影响。这项工作将通过允许有界违反数据库的完整性约束来考虑可序列化和吞吐量/响应时间之间的权衡。当冲突的事务不同步时，就会发生此类违规行为。在这些情况下，事务不知道先前某些事务的影响，因此它返回的结果和对数据库的修改都是基于不完全信息。与可序列化行为的偏差和吞吐量/响应时间都与忽略的程度直接相关。并发控制算法限制了事务可以忽略冲突事务的程度，将限制与可串行化行为的偏差。由于完整性约束的违反并不是在所有应用中都是允许的，因此该研究涉及到对其中可以容忍违反的应用的特征的研究。

The objective of this research is to develop concurrency control algorithms for transactions in distributed database systems which achieve high throughput and low response time. Data replication is a mechanism that has been proposed for this purpose. Unfortunately, the requirement that transactions behave serializably forces considerable synchronization among replica sites and hence adversely impacts throughput and response time. This work will consider the tradeoff between serializability and throughput/response time by allowing bounded violations of the database's integrity constraints. Such violations occur when conflicting transactions are not synchronized. Under these circumstances a transaction is ignorant of the effects of some prior transactions, and so the results it returns and the modifications it makes to the database are based on incomplete information. Both the deviation from serializable behavior and the throughput/response time are directly related to the extent of ignorance. Concurrency control algorithms which bound the extent to which a transaction can be ignorant of conflicting transaction will bound the deviation from serializable behavior. Since violations of integrity constraints are not permissible in all applications, the research involves study of the characteristics of applications in which violations can be tolerated.