Dynamic Multi-Resource Load Balancing in Parallel Database Systems
Erhard Rahm, Robert Marek · Qucosa (Saxon State and University Library Dresden) · 1995
Parallel database systems have to support the effective parallelization of complex queries in multi-user mode, i.e. in combination with inter-query~mter-transaction parallelism. For this purpose, dynamic scheduling and load balancing strate-DBMS applications increasingly face the need of parallel query processing due to growing database sixes and query complexity. In addition, high trausaction rates must be supported for standard OLTP applications. gies ’ are necessary that umsider the current system state for The effective use of super-serve & for database processing dekrminhg the degree of intra-query parallelism and for se- poses many implementation challenges that am largely unlecting the processors for executing subqueries. We study solved in current products 128.113. One key problem is the these issues for parallel hash joinprocessing and show that effective use of inuaquery parallelism in multi-user mode, the two subproblems should be addressed in au integrated i.e., when complex queries am executed concurrently with way. Even more importantly, however, is the use of a multi- OLTP transactions and other complex que&s. Multi-user mannce load balancing approach that considers all potential mode (inter-trausactio&mter-query parallelism) is mandatobottleneck resources. in particular memory, disk and CPU. ry to achieve acceptable throughput and cost-effectiveness, We discuss basic performance tradeoffs to consider and eval- in particular for super-servers wherea high numberof pro-Gate the performauce of several load balancing strategies by cessors must efgectively be utilii. While proposed algomeans of a detailed simulation model. Simulation results will rithms for parallel query processing also work in multi-user be analyzed for multiuser configurations with both homoge- mode. their perfcrmance may be substantially lower than in neous andheterogeneous (query/OLTP) workloads. single-user mode. This is because multi-user mode inevitably leads to data and resource contention that can significant-1