Efficient Sorting and Join on NVM-Based Hybrid Memory

Abstract

Non-volatile memory (NVM) as a new kind of future memories has a number of special properties such as non-volatility, read/write asymmetry, and byte addressability. This makes it difficult to directly replace DRAM with NVM in current memory hierarchy. Thus, a practical way is to construct hybrid memory composed of both NVM and DRAM. Such hybrid memory architecture introduces many new challenges for existing algorithms. In this paper, we focus on improving sorting and join algorithms for DRAM-NVM-based hybrid memory. In particular, we start with a theoretical study on the data placement issue in DRAM-NVM-based hybrid memory systems and propose an optimal data placement model to store data structures in DRAM and NVM during the sorting process. We present the theoretical proof to the optimal data placement model to ensure the correctness of the model. Further, based on the optimal data placement model, we propose a new NVM-aware sorting algorithm named NVMSort that adopts heap structures to accelerate the sorting process. Compared with traditional sorting algorithms, NVMSort is write-friendly and more efficient on DRAM-NVM-based hybrid memory. We further apply NVMSort into the traditional merge-sort join algorithm to optimize merge-sort join on DRAM-NVM-based hybrid memory. We conduct comparative experiments with existing sorting algorithms including HeapSort and QuickSort. The results show that NVMSort is much faster than the classical Heapsort and QuickSort. In addition, NVMSort is more NVM-friendly as it can reduce more NVM writes. When integrated into the traditional merge-sort join algorithm, NVMSort also achieves the best performance.