Dynamic load balancing is a prerequisite for effectively utilizing large parallel database systems. Load balancing at different levels is required in particular for assigning transactions and queries as well as subqueries to nodes. Special problems are posed by the need to support both inter-transaction/query as well as intra-transaction/query parallelism due to conflicting performance requirements. We compare the major architectures for parallel database systems, Shared Nothing and Shared Disk, with respect to their load balancing potential. For this purpose, we focus on parallel scan and join processing in multi-user mode. It turns out that both the degree of query parallelism as well as the processor allocation should be determined in a coordinated way and based on the current utilization of critical resource types, in particular CPU and memory.