A Preliminary Study of Scalability of TCP/IP Based Clusters under Database Workloads

摘要

In this paper we study the scalability of non-partitioned, clustered database management systems as a function of inter-process communication (IPC) latency and number of nodes. It is assumed that the clustered DBMS has a fully shared IO subsystem and multiversion concurrency control over the data in various buffer caches. The cluster interconnect fabric is assumed to be TCP/IP over Ethernet with and w/o hardware offload. The main contribution of the paper is to shed some light on the scalability of DBMS workloads in a scaleout environment as a function of number of nodes and interconnect latencies.