Parallel Algorithms for Top-k Query Processing

The general problem of answering top-k queries can be modeled using lists of objects sorted by their local scores. Fagin et al. proposed the “middleware cost” for a top-k query algorithm, and proposed the efficient sequential Threshold Algorithm (TA). However, since the size of the dataset can be incredible huge, the middleware cost of sequential TA may be intolerable. So, in this paper, we propose parallel algorithms to process top-k queries and analyze their middleware costs. Intuitively, a naive parallel algorithm, called PTA (parallel-TA), evenly partitions the original dataset into P (the number of processors) subdatasets. Each processor finds top-k results of one corresponding subdataset using TA algorithm. Then these results are merged to get the final top-k answers. Motivated by the idea of partitioning objects, we take a further step to partition D into n subdatasets according to their degree of domination. Based on this partition, we propose EPTA (Enhanced-PTA) algorithm. Under PRAM-CRCW model, the middleware cost of PTA is 2 ( / ) O nm P while the average middleware cost of EPTA is 2 -1 ( (ln ) / ( -1)!) m O km n m under the assumption that scores in different lists are independently distributed, where n is the dataset size and m is the number of lists. Extensive experiments show that the speedup ratios of EPTA are significantly higher than those of PTA.