Issues facing commercial OLTP applications on MPP platforms

The author examines the scalability of massively parallel processing (MPP) architectures executing online transaction processing (OLTP) workloads on commercially available relational database systems (RDBMS). Most MPP hardware architectures are either shared nothing or shared memory. In the near future, the available scalability will be impacted primarily by two major influences: commodity hardware performance and the performance of the RDBMS architecture. The author presents a simple scaling model for MPPs running such workloads. He introduces the new concept of super-seriality to account for potential MPP performance penalties. Super-seriality has the effect of inducing a premature maximum in the scaling curve beyond which it is counter-productive to add further processors. This results in unexpected limitations in the scalability of MPPs executing OLTP workloads. Scaleup predictions are presented for some existing transaction processing platforms.<>