Shared-nothing并行数据库系统查询优化技术

Query Optimization Techniques of A Shared nothing Parallel Database System

查询优化是并行数据库系统的核心技术 .该文介绍作者自行研制的一个 Shared- nothing并行数据库系统PBA SE/ 2中独特的两阶段优化策略 .为了缩减并行查询优化庞大的搜索空间 ,PBASE/ 2将并行查询优化划分为顺序优化和并行化两个阶段 .在顺序优化阶段对并行化后的通信代价进行预先估算 ,将通信开销加入顺序优化的代价模型 ,同时对动态规划搜索算法进行了修正和扩展 ,保证了顺序优化阶段得到的最小代价计划在并行化后代价仍然最小 .并行化阶段的优化目标是实现查询工作量在系统内多种资源上的负载平衡 ,提出了资源负载平衡因子的概念 ,并且通过启发式规则、任务调度等机制保证了并行查询执行计划的优化性 .

Query optimization is the critical technique of parallel database systems. This paper addresses the distinguishing two phase optimization strategy implemented in a shared nothing parallel database system PBASE/2. In order to cut down the enormous search space of parallel query optimization, PBASE/2 divides parallel query processing into two phases, sequential optimization and parallelization. In sequential optimization, PBASE/2 predicts the communication cost of parallel execution, and addes communication cost into cost model. At the same time, the traditional dynamic programming search algorithm is modified and extended. By these mechanisms, it is guaranteed that the sequential minimal cost plan also is the parallel minimal cost plan. The object of parallelization phase is to balance the query work loading among various resources in the system. Thus, a new concept, resources load balancing factor, is put forward. Other techniques such as heuristics, dynamic programming based parallelizaion algorithm, and task scheduling ensure the quality of the outgoing parallel query evaluation plans.