Chopping Up Trees to Improve Spatial Locality in Implicit k-Heaps

Research on the performance of implicit k-heaps has shown that aligning data with cache lines and increasing heap arity are effective techniques for improving the data reference locality of heap operations. The technique of tree blocking has long been used to enhance the data reference locality of tree-based search methods. In this paper we propose c-clustered tree blocking, a new tree blocking method designed to further enhance the data reference locality of implicit k-heap operations. We examine the effect of our method on the performance of a traditional aligned implicit 2-heap using internal memory benchmarks based on the Hold model. Our empirical results, reproduced on four contemporary architectures, show that our method produces speedups of up to 2.0 in either benchmark, while reducing data cache misses by up to 85% and TLB misses by up to 65%. For larger heap arities our method matches the performance of traditional implicit k-heaps while improving page level locality.