CPU and GPU Hash Joins on Skewed Data

Hash join is one of the most important and widely used query processing operations. In many real-world applications, such as graph data analysis, join keys can be highly skewed. However, data skew is often only a secondary design consideration in existing CPU and GPU hash join algorithms. We find that the data structures and algorithm steps in existing hash joins are less efficient for handling highly skewed join keys, leading to significant performance degradation. In this paper, we optimize hash joins for skewed data, and propose a CPU Skewconscious Hash join (CSH) and a GPU Skew-conscious Hash join (GSH). Our main idea is to detect skewed join keys and handle skewed vs. normal keys in separate routines optimized for their target cases. Our preliminary experiments show that compared to state-of-the-art CPU and GPU hash joins with existing skew-handling techniques, our proposed CSH and GSH achieve up to 8.0x and 13.5x improvement, respectively.