I. Introduction
Recently, data mining is a highlighted technology to discover the valuable trends behind data [1]. In the data mining area, frequent-pattern mining is a key problem that identifies the frequent-occurring itemsets or patterns in a given dataset. While many frequent-pattern mining methods like Apriori [2] and frequent-pattern tree (FP-tree) [3] have been widely used, they are primarily designed for volatile and energy-inefficient dynamic random-access memory (DRAM). To enhance the persistence and energy efficiency of frequent-pattern mining, modern nonvolatile memories (NVMs), such as phase-change memory (PCM) [4] are considerable alternatives to DRAM in diversified computing systems like embedded systems to keep the mining metadata, so as to facilitate the high-performance in-memory data analytics. Unfortunately, the distinct characteristics of NVMs, such as skewed write performance and energy [4], [5], might degrade the performance and energy efficiency of existing mining methods on NVMs. Although the problem could be alleviated by jointly concerning the NVM characteristics into the design of the frequent-pattern mining methods [6], the scalability of the mining methods is still limited where the to-be-mined dataset is gigantic and the number of distinct data items is huge. This article is therefore motivated by proposing a highly scalable method for the in-memory frequent-pattern mining problem. To be specific, we augment an existing frequent-pattern mining method, i.e., the popular FP-tree approach, to utilize the performance benefits of symmetric multiprocessor (SMP) parallel computing architecture [7]. Meanwhile, the NVM characteristics are jointly considered into the design of the augmented method, so as to alleviate the undesirable degradation of the performance and energy efficiency of frequent-pattern mining.