Dynamic memory allocation strategies for parallel query execution

In the decision support queries which manipulate large data volumes, it is frequent that a query constituted by several joins can not be computed completely in memory. In this paper, we propose three strategies allowing to assign the memory of a shared-nothing parallel architecture to operation clones of a query. The performance evaluation of the three strategies shows that the strategies which favor the operation clones using a lot of memory obtain a better response time than the strategy which favors the clones using little memory. The main contribution of this paper is to take into account the available memory sizes on every processor and to avoid allotting the same processor to two operation clones that must run in parallel.