Clustered DBMS scalability under unified Ethernet fabric

In this paper, we study the performance of a clustered DBMS running an on-line transaction processing (OLTP) workload and using TCP/IP over Ethernet as a "unified fabric" for inter-process communication. The study is based on a comprehensive simulation model of such systems, called DCLUE. It is found that while the protocol overhead has a large impact on performance, the end to end latency has significantly lesser impact. The results suggest that very low-latency protocol offload may not be very beneficial for clustered OLTP workloads, but proper QoS treatment of messages is important.