Scalable Parallel and Distributed Expert Database Systems

我们使用PARADISER 10分布式规则处理环境作为我们的测试平台，研究专家数据库中性能瓶颈的解决方案。我们的目的是采用一种方法，结合对规则程序规则的静态计算限制，将规则评估的工作负载划分为在不同处理站点评估的任意数量的规则程序副本，以及动态负载平衡协议，该协议通过在运行时修改限制来更新和重组工作负载的分布。（规则不会重新分配。）我们的技术利用了一种形式的“元数据”，即在计算规则限制的算法中收集属性值和底层数据库关系大小的统计数据（例如，属性的不同值的数量，关系中的元组的数量等）。我们从效率和可伸缩性标准方面分析了动态负载平衡协议，这为随着数据库大小的增加，协议的预期行为提供了指导。提出的研究分为三个主要类别：(a)处理站点之间的工作负载分配范式，以改善因数据库规模扩大而导致的性能下降；(b)一套用于底层数据库动态“逻辑”重组的负载平衡协议及其扩展特性；(c)在PARADISER专家数据库环境中使用一套基准规则程序进行实验验证。将构建一个完全可操作且可转移的规则处理环境来体现这些结果。

9313847 Stolfo We study solutions to the performance bottleneck in expert databases using the PARADISER 10 distributed rule processing environment as our testbed. We purpose an approach that combines statically computed restrictions on the rules of a rule program for partitioning the workload of rule evaluation among an arbitrary number of rule program replicas evaluated at distinct processing sites, and dynamic load balancing protocols that update and reorganize the distribution of workload by modifying the restrictions at runtime. (Rules are not redistributed.) Our techniques utilize one form of "meta-data", namely, statistics gathered on the attribute values and size of relations of the underlying database (e.g., number of distinct values of an attribute, number of tuples in a relation etc.) in the algorithms that compute the restrictions on the rules. We analyze the dynamic load balancing protocols in terms of efficiency and scalability criteria, which provide guidance on the expected behavior of the protocols with increasing database size. The proposed research falls into three main categories: (a) workload distribution paradigms among processing sites to ameliorate the performance degradation resulting from scaling up of database size, (b) a suite of load balancing protocols for dynamic "logical" reorganization of the underlying database, and their scaling characteristics, and , (c) experimental validation in the PARADISER expert database environment, using a suite of benchmark rule programs. A fully operational and transferable rule processing environment will be constructed that embodies these results.