Analysis of Dynamic Load Balancing Strategies for Parallel Shared Nothing Database Systems

Parallel database systems have to support both inter-transaction as well as intra-transaction parallelism. Inter-transaction parallelism (multi-user mode) is required to achieve high throughput, in particular for OLTP transactions, and sufficient cost-effectiveness. Intra-transaction parallelism is a prerequisite for reducing the response time of complex and data-intensive transactions (queries). In order to achieve both goals dynamic strategies for load balancing and scheduling are necessary which take the current system state into account for allocating transactions and subqueries to processors and for determining the degree of intra-transaction parallelism. We study the load balancing problem for parallel join processing in Shared Nothing database systems. In these systems, join processing is typically based on a dynamic redistribution of relations to join processors thus making dynamic load balancing strategies feasible. In particular, we study the performance of dynamic load balancing strategies for determining the number of join processors and for selection of the join processors. In contrast to previous studies on parallel join processing, we present a multi-user performance analysis for both homogeneous and heterogeneous/mixed workloads as well as for different database allocations.