Title:CARMI: A Cache-Aware Learned Index with a Cost-based Construction Algorithm

Abstract:Learned indexes, which use machine learning models to replace traditional index structures, have shown promising results in recent studies. However, our understanding of this new type of index structure is still at an early stage with many details that need to be carefully examined and improved. In this paper, we propose a cache-aware learned index (CARMI) design to improve the efficiency of the Recursive Model Index (RMI) framework proposed by Kraska et al. and a cost-based construction algorithm to construct the optimal indexes in a wide variety of application scenarios. We formulate the problem of finding the optimal design of a learned index as an optimization problem and propose a dynamic programming algorithm for solving it and a partial greedy step to speed up. Experiments show that our index construction strategy can construct indexes with significantly better performance compared to baselines under various data distribution and workload requirements. Among them, CARMI can obtain an average of 2.52X speedup compared to B-tree, while using only about 0.56X memory space of B-tree on average.