Consideration of Page-Splitting Methods in a Fat-Btree for Distributing PostgreSQL
Yuta Namiki, Kota Kanbe, D. Kobayashi, Haruo Yokota · IEICE Technical Report; IEICE Tech. Rep. · 2007
We are trying to construct a shared-nothing parallel DBMS by storing data in a Fat-Btree, a parallel B-tree structure capable of dynamically changing data distribution. For page splits caused by inserting tuples in the Fat-Btree, we have proposed two types of methods to cooperate with the MVCC of PostgreSQL: One splits a page when it becomes full, while the other preparatory does when its occupation rate exceeds a threshold. In this paper, we implement the system which distribute PostgreSQL by using the Fat-Btree, and evaluate two page-splitting methods. Experimental results show that the preparatory-split method provides higher throughput than the other in frequent update requests.