Kevin H. Liu

Performance evaluation of processor allocation algorithms for parallel query execution

Two processor allocation algorithms are presented based on a query cost model incorporating the effects of data cormnunication overheads and load imbalance running on a shared nothing parallel architecture. The phase-based algorithm makes use of the heuristic of merge-point evaluation so that the number of operations in each execution phase are distributed evenly. Time equalisation technique is employed within each phase to minimise phase execution time. The non phase-based algorithm is a dynamic approach and its performance is sensitive to the number of processors available. Basea on the algorithm, the operations in the query tree are c :ecuted in such a order that leaf operations are conducted first and processing is ended with the root operation. The concept of optimal degree of parallelism for each operation is also introduced and the new algorithms are evaluated with a simulation study. The experimental results show that the new algorithms outperform the conventional processor allocation algorithms. In addition, the phase based algorithm always provides a local phase optimisation and the non phase based algorithm gives a global optimal solution when there is a large number of processors available.

Performance study on optimal processor assignment in parallel relational databases

In this paper, we introduce the concept of optimal degree of parallelism into both individual operation and single query level, and present a subtree-grouping method, SCM, to search the optimal processor assignment. By making use of the introduced concept and SCM, an optimised processor assignment algorithm, OPA, is developed by employing both intra and inter operation parallelism. OPA provides a global optimal solution when the number of processors is sufficient and a local phase optimisation in all other situations. The new algorithm is able to perform adaptively based on two methods; dynamic processor assignment method where the optimal number of processors are allocated to operations during processing, and merge-point-oriented phase partitioning method where operations are grouped into execution phases evenly based on the heuristic of merge point evaluation. The new algorithm is evaluated by a simulation study together with the existing algorithms. The results show that the new algorithm always gives the best performance based on a large number of queries in five different query groups.

Multiple dependent queries execution using critical path scheduling in parallel databases

Multiple processors are employed to improve the performance of database systems and the parallelism can be exploited at three levels in query processing: intra-operation, inter-operation, and inter-query parallelism. Intra-operation and inter-operation parallelism are also called intra-query parallelism which has been studied extensively. In contrast, inter-query parallelism has received little attention particularly for multiple dependent queries. We develop a decompression algorithm, CPS, for coping with multiple dependent queries which are represented by a directed graph, and the algorithm makes use of the activity analysis of critical path analysis, and the resource scheduling and levelling of project management. A simulation study has been conducted and the results show that the proposed algorithm outperforms other existing methods and is able to provide a global optimal solution when the number of processors available is sufficient.